[OpenMP] Change OpenMP code generation for target region entries

This patch changes the code we generate to enter a target region on the
device. This is in-line with the new definition in the runtime that was
added previously. Additionally we implement this in the OpenMPIRBuilder
so that this code can be shared with Flang in the future.

Reviewed By: ABataev

Differential Revision: https://reviews.llvm.org/D128550

diff --git a/clang/lib/CodeGen/CGOpenMPRuntime.cpp b/clang/lib/CodeGen/CGOpenMPRuntime.cpp
index 305040b..e412e39 100644 (file)
--- a/clang/lib/CodeGen/CGOpenMPRuntime.cpp
+++ b/clang/lib/CodeGen/CGOpenMPRuntime.cpp
@@ -6717,11 +6717,9 @@ llvm::Value *CGOpenMPRuntime::emitNumTeamsForTargetDirective(
     default:
       break;
     }
-  } else if (DefaultNT == -1) {
-    return nullptr;
   }
 
-  return Bld.getInt32(DefaultNT);
+  return llvm::ConstantInt::get(CGF.Int32Ty, DefaultNT);
 }
 
 static llvm::Value *getNumThreads(CodeGenFunction &CGF, const CapturedStmt *CS,
@@ -10311,23 +10309,29 @@ void CGOpenMPRuntime::emitTargetCall(
     // Emit tripcount for the target loop-based directive.
     emitTargetNumIterationsCall(CGF, D, DeviceID, SizeEmitter);
 
-    bool HasNowait = D.hasClausesOfKind<OMPNowaitClause>();
+    // Arguments for the target kernel.
+    SmallVector<llvm::Value *> KernelArgs{
+        CGF.Builder.getInt32(/* Version */ 1),
+        PointerNum,
+        InputInfo.BasePointersArray.getPointer(),
+        InputInfo.PointersArray.getPointer(),
+        InputInfo.SizesArray.getPointer(),
+        MapTypesArray,
+        MapNamesArray,
+        InputInfo.MappersArray.getPointer()};
+
+    // Arguments passed to the 'nowait' variant.
+    SmallVector<llvm::Value *> NoWaitKernelArgs{
+        CGF.Builder.getInt32(0),
+        llvm::ConstantPointerNull::get(CGM.VoidPtrTy),
+        CGF.Builder.getInt32(0),
+        llvm::ConstantPointerNull::get(CGM.VoidPtrTy),
+    };
+
+    bool HasNoWait = D.hasClausesOfKind<OMPNowaitClause>();
+
     // The target region is an outlined function launched by the runtime
-    // via calls __tgt_target() or __tgt_target_teams().
-    //
-    // __tgt_target() launches a target region with one team and one thread,
-    // executing a serial region.  This master thread may in turn launch
-    // more threads within its team upon encountering a parallel region,
-    // however, no additional teams can be launched on the device.
-    //
-    // __tgt_target_teams() launches a target region with one or more teams,
-    // each with one or more threads.  This call is required for target
-    // constructs such as:
-    //  'target teams'
-    //  'target' / 'teams'
-    //  'target teams distribute parallel for'
-    //  'target parallel'
-    // and so on.
+    // via calls to __tgt_target_kernel().
     //
     // Note that on the host and CPU targets, the runtime implementation of
     // these calls simply call the outlined function without forking threads.
@@ -10338,70 +10342,15 @@ void CGOpenMPRuntime::emitTargetCall(
     // In contrast, on the NVPTX target, the implementation of
     // __tgt_target_teams() launches a GPU kernel with the requested number
     // of teams and threads so no additional calls to the runtime are required.
-    if (NumTeams) {
-      // If we have NumTeams defined this means that we have an enclosed teams
-      // region. Therefore we also expect to have NumThreads defined. These two
-      // values should be defined in the presence of a teams directive,
-      // regardless of having any clauses associated. If the user is using teams
-      // but no clauses, these two values will be the default that should be
-      // passed to the runtime library - a 32-bit integer with the value zero.
-      assert(NumThreads && "Thread limit expression should be available along "
-                           "with number of teams.");
-      SmallVector<llvm::Value *> OffloadingArgs = {
-          RTLoc,
-          DeviceID,
-          OutlinedFnID,
-          PointerNum,
-          InputInfo.BasePointersArray.getPointer(),
-          InputInfo.PointersArray.getPointer(),
-          InputInfo.SizesArray.getPointer(),
-          MapTypesArray,
-          MapNamesArray,
-          InputInfo.MappersArray.getPointer(),
-          NumTeams,
-          NumThreads};
-      if (HasNowait) {
-        // Add int32_t depNum = 0, void *depList = nullptr, int32_t
-        // noAliasDepNum = 0, void *noAliasDepList = nullptr.
-        OffloadingArgs.push_back(CGF.Builder.getInt32(0));
-        OffloadingArgs.push_back(llvm::ConstantPointerNull::get(CGM.VoidPtrTy));
-        OffloadingArgs.push_back(CGF.Builder.getInt32(0));
-        OffloadingArgs.push_back(llvm::ConstantPointerNull::get(CGM.VoidPtrTy));
-      }
-      Return = CGF.EmitRuntimeCall(
-          OMPBuilder.getOrCreateRuntimeFunction(
-              CGM.getModule(), HasNowait
-                                   ? OMPRTL___tgt_target_teams_nowait_mapper
-                                   : OMPRTL___tgt_target_teams_mapper),
-          OffloadingArgs);
-    } else {
-      SmallVector<llvm::Value *> OffloadingArgs = {
-          RTLoc,
-          DeviceID,
-          OutlinedFnID,
-          PointerNum,
-          InputInfo.BasePointersArray.getPointer(),
-          InputInfo.PointersArray.getPointer(),
-          InputInfo.SizesArray.getPointer(),
-          MapTypesArray,
-          MapNamesArray,
-          InputInfo.MappersArray.getPointer()};
-      if (HasNowait) {
-        // Add int32_t depNum = 0, void *depList = nullptr, int32_t
-        // noAliasDepNum = 0, void *noAliasDepList = nullptr.
-        OffloadingArgs.push_back(CGF.Builder.getInt32(0));
-        OffloadingArgs.push_back(llvm::ConstantPointerNull::get(CGM.VoidPtrTy));
-        OffloadingArgs.push_back(CGF.Builder.getInt32(0));
-        OffloadingArgs.push_back(llvm::ConstantPointerNull::get(CGM.VoidPtrTy));
-      }
-      Return = CGF.EmitRuntimeCall(
-          OMPBuilder.getOrCreateRuntimeFunction(
-              CGM.getModule(), HasNowait ? OMPRTL___tgt_target_nowait_mapper
-                                         : OMPRTL___tgt_target_mapper),
-          OffloadingArgs);
-    }
-
     // Check the error code and execute the host version if required.
+    CGF.Builder.restoreIP(
+        HasNoWait ? OMPBuilder.emitTargetKernel(
+                        CGF.Builder, Return, RTLoc, DeviceID, NumTeams,
+                        NumThreads, OutlinedFnID, KernelArgs, NoWaitKernelArgs)
+                  : OMPBuilder.emitTargetKernel(CGF.Builder, Return, RTLoc,
+                                                DeviceID, NumTeams, NumThreads,
+                                                OutlinedFnID, KernelArgs));
+
     llvm::BasicBlock *OffloadFailedBlock =
         CGF.createBasicBlock("omp_offload.failed");
     llvm::BasicBlock *OffloadContBlock =

diff --git a/clang/test/OpenMP/capturing_in_templates.cpp b/clang/test/OpenMP/capturing_in_templates.cpp
index 9423996..f270775 100644 (file)
--- a/clang/test/OpenMP/capturing_in_templates.cpp
+++ b/clang/test/OpenMP/capturing_in_templates.cpp
@@ -24,7 +24,7 @@ pair<T1, T2> make_pair(T1 &&t1, T2 &&t2) {
 
 // CHECK-LABEL: @main
 int main(int argc, char **argv) {
-// CHECK: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null)
+// CHECK: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 #pragma omp target
  {
     for (int i = 0; i < 64; ++i) {

diff --git a/clang/test/OpenMP/declare_mapper_codegen.cpp b/clang/test/OpenMP/declare_mapper_codegen.cpp
index 0c8c411..aef2fcd 100644 (file)
--- a/clang/test/OpenMP/declare_mapper_codegen.cpp
+++ b/clang/test/OpenMP/declare_mapper_codegen.cpp
@@ -243,20 +243,27 @@ void foo(int a){
   C c;
   c.a = a;
 
-  // CK0-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** [[MPRGEP:%.+]])
-  // CK0-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK0-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK0-DAG: [[MPRGEP]] = bitcast [1 x i8*]* [[MPR:%[^,]+]] to i8**
-  // CK0-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK0-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK0-DAG: [[MPR1:%.+]] = getelementptr inbounds {{.+}}[[MPR]], i[[sz]] 0, i[[sz]] 0
-  // CK0-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to %class.C**
-  // CK0-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to %class.C**
-  // CK0-DAG: store %class.C* [[VAL:%[^,]+]], %class.C** [[CBP1]]
-  // CK0-DAG: store %class.C* [[VAL]], %class.C** [[CP1]]
-  // CK0-DAG: store i8* bitcast (void (i8*, i8*, i8*, i64, i64, i8*)* [[MPRFUNC]] to i8*), i8** [[MPR1]]
-  // CK0: call void [[KERNEL_1:@.+]](%class.C* [[VAL]])
-  #pragma omp target map(mapper(id),tofrom: c)
+// CK0-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK0-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK0-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK0-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK0-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK0-DAG: [[MARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 7
+// CK0-DAG: store i8** [[MPRGEP:%.+]], i8*** [[MARG]]
+// CK0-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK0-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK0-DAG: [[MPRGEP]] = bitcast [1 x i8*]* [[MPR:%[^,]+]] to i8**
+// CK0-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK0-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK0-DAG: [[MPR1:%.+]] = getelementptr inbounds {{.+}}[[MPR]], i[[sz]] 0, i[[sz]] 0
+// CK0-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to %class.C**
+// CK0-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to %class.C**
+// CK0-DAG: store %class.C* [[VAL:%[^,]+]], %class.C** [[CBP1]]
+// CK0-DAG: store %class.C* [[VAL]], %class.C** [[CP1]]
+// CK0-DAG: store i8* bitcast (void (i8*, i8*, i8*, i64, i64, i8*)* [[MPRFUNC]] to i8*), i8** [[MPR1]]
+// CK0: call void [[KERNEL_1:@.+]](%class.C* [[VAL]])
+#pragma omp target map(mapper(id), tofrom \
+                       : c)
   {
     ++c.a;
   }
@@ -282,20 +289,27 @@ void foo(int a){
     ++c.a;
   }
 
-  // CK0-DAG: call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[TEAMSIZES]]{{.+}}, {{.+}}[[TEAMTYPES]]{{.+}}, i8** null, i8** [[MPRGEP:%.+]], i32 0, i32 0)
-  // CK0-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK0-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK0-DAG: [[MPRGEP]] = bitcast [1 x i8*]* [[MPR:%[^,]+]] to i8**
-  // CK0-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK0-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK0-DAG: [[MPR1:%.+]] = getelementptr inbounds {{.+}}[[MPR]], i[[sz]] 0, i[[sz]] 0
-  // CK0-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to %class.C**
-  // CK0-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to %class.C**
-  // CK0-DAG: store %class.C* [[VAL:%[^,]+]], %class.C** [[CBP1]]
-  // CK0-DAG: store %class.C* [[VAL]], %class.C** [[CP1]]
-  // CK0-DAG: store i8* bitcast (void (i8*, i8*, i8*, i64, i64, i8*)* [[MPRFUNC]] to i8*), i8** [[MPR1]]
-  // CK0: call void [[KERNEL_3:@.+]](%class.C* [[VAL]])
-  #pragma omp target teams map(mapper(id),to: c)
+// CK0-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 0, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK0-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK0-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK0-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK0-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK0-DAG: [[MARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 7
+// CK0-DAG: store i8** [[MPRGEP:%.+]], i8*** [[MARG]]
+// CK0-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK0-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK0-DAG: [[MPRGEP]] = bitcast [1 x i8*]* [[MPR:%[^,]+]] to i8**
+// CK0-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK0-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK0-DAG: [[MPR1:%.+]] = getelementptr inbounds {{.+}}[[MPR]], i[[sz]] 0, i[[sz]] 0
+// CK0-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to %class.C**
+// CK0-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to %class.C**
+// CK0-DAG: store %class.C* [[VAL:%[^,]+]], %class.C** [[CBP1]]
+// CK0-DAG: store %class.C* [[VAL]], %class.C** [[CP1]]
+// CK0-DAG: store i8* bitcast (void (i8*, i8*, i8*, i64, i64, i8*)* [[MPRFUNC]] to i8*), i8** [[MPR1]]
+// CK0: call void [[KERNEL_3:@.+]](%class.C* [[VAL]])
+#pragma omp target teams map(mapper(id), to \
+                             : c)
   {
     ++c.a;
   }
@@ -493,7 +507,15 @@ void foo(int a){
 // CK0: }
 
 // CK0: define internal void [[OUTLINED:@.+]](i32 {{.*}}{{[^,]+}}, [[ANON_T]]* noalias noundef [[CTXARG:%.+]])
-// CK0-DAG: call i32 @__tgt_target_nowait_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZEGEP:%[0-9]+]], {{.+}}[[NWTYPES]]{{.+}}, i8** null, i8** [[MPRGEP:%.+]], i32 0, i8* null, i32 0, i8* null)
+// CK0-DAG: call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]], i32 0, i8* null, i32 0, i8* null)
+// CK0-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK0-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK0-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK0-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK0-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK0-DAG: store i64* [[SIZEGEP:%.+]], i64** [[SARG]]
+// CK0-DAG: [[MARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 7
+// CK0-DAG: store i8** [[MPRGEP:%.+]], i8*** [[MARG]]
 // CK0-DAG: [[BPGEP]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[BPFPADDR:%.+]], i[[SZ]] 0, i[[SZ]] 0
 // CK0-DAG: [[PGEP]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[PFPADDR:%.+]], i[[SZ]] 0, i[[SZ]] 0
 // CK0-DAG: [[SIZEGEP]] = getelementptr inbounds [1 x i64], [1 x i64]* [[SIZEFPADDR:%.+]], i[[SZ]] 0, i[[SZ]] 0
@@ -525,7 +547,15 @@ void foo(int a){
 // CK0: }
 
 // CK0: define internal void [[OUTLINE_1:@.+]](i32 {{.*}}%.global_tid.{{.+}}, [[ANON_T_0]]* noalias noundef [[CTXARG:%.+]])
-// CK0-DAG: call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], i64* [[SIZEGEP:%[0-9]+]], {{.+}}[[TEAMNWTYPES]]{{.+}}, i8** null, i8** [[MPRGEP:%.+]], i32 0, i32 0, i32 0, i8* null, i32 0, i8* null)
+// CK0-DAG: call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 -1, i32 0, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]], i32 0, i8* null, i32 0, i8* null)
+// CK0-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK0-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK0-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK0-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK0-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK0-DAG: store i64* [[SIZEGEP:%.+]], i64** [[SARG]]
+// CK0-DAG: [[MARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 7
+// CK0-DAG: store i8** [[MPRGEP:%.+]], i8*** [[MARG]]
 // CK0-DAG: [[BPGEP]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[BPFPADDR:%.+]], i[[SZ]] 0, i[[SZ]] 0
 // CK0-DAG: [[PGEP]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[PFPADDR:%.+]], i[[SZ]] 0, i[[SZ]] 0
 // CK0-DAG: [[SIZEGEP]] = getelementptr inbounds [1 x i64], [1 x i64]* [[SIZEFPADDR:%.+]], i[[SZ]] 0, i[[SZ]] 0
@@ -902,29 +932,36 @@ void foo(int a){
 
   // CK3-DAG: [[BC:%.+]] = getelementptr inbounds %class.B, %class.B* [[BVAL]], i32 0, i32 0
 
-  // CK3-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 2, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** [[MPRGEP:%.+]])
-  // CK3-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK3-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK3-DAG: [[MPRGEP]] = bitcast [2 x i8*]* [[MPR:%[^,]+]] to i8**
-  // CK3-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK3-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK3-DAG: [[MPR1:%.+]] = getelementptr inbounds {{.+}}[[MPR]], i{{64|32}} 0, i{{64|32}} 0
-  // CK3-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to %class.B**
-  // CK3-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to %class.C**
-  // CK3-DAG: store %class.B* [[BVAL]], %class.B** [[CBP1]]
-  // CK3-DAG: store %class.C* [[BC]], %class.C** [[CP1]]
-  // CK3-DAG: store i8* bitcast (void (i8*, i8*, i8*, i64, i64, i8*)* [[MPRFUNC]] to i8*), i8** [[MPR1]]
-  // CK3-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
-  // CK3-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
-  // CK3-DAG: [[MPR2:%.+]] = getelementptr inbounds {{.+}}[[MPR]], i{{64|32}} 0, i{{64|32}} 1
-  // CK3-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [10 x %class.C]**
-  // CK3-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to %class.C**
-  // CK3-DAG: store [10 x %class.C]* [[CVAL]], [10 x %class.C]** [[CBP2]]
-  // CK3-DAG: [[CVALGEP:%.+]] = getelementptr inbounds {{.+}}[[CVAL]], i{{64|32}} 0, i{{64|32}} 0
-  // CK3-DAG: store %class.C* [[CVALGEP]], %class.C** [[CP2]]
-  // CK3-DAG: store i8* bitcast (void (i8*, i8*, i8*, i64, i64, i8*)* [[MPRFUNC]] to i8*), i8** [[MPR2]]
-  // CK3: call void [[KERNEL:@.+]](%class.B* [[BVAL]], [10 x %class.C]* [[CVAL]])
-  #pragma omp target map(mapper(id),tofrom: c[0:10], b.c)
+// CK3-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK3-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK3-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK3-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK3-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK3-DAG: [[MARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 7
+// CK3-DAG: store i8** [[MPRGEP:%.+]], i8*** [[MARG]]
+// CK3-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK3-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK3-DAG: [[MPRGEP]] = bitcast [2 x i8*]* [[MPR:%[^,]+]] to i8**
+// CK3-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK3-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK3-DAG: [[MPR1:%.+]] = getelementptr inbounds {{.+}}[[MPR]], i{{64|32}} 0, i{{64|32}} 0
+// CK3-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to %class.B**
+// CK3-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to %class.C**
+// CK3-DAG: store %class.B* [[BVAL]], %class.B** [[CBP1]]
+// CK3-DAG: store %class.C* [[BC]], %class.C** [[CP1]]
+// CK3-DAG: store i8* bitcast (void (i8*, i8*, i8*, i64, i64, i8*)* [[MPRFUNC]] to i8*), i8** [[MPR1]]
+// CK3-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
+// CK3-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
+// CK3-DAG: [[MPR2:%.+]] = getelementptr inbounds {{.+}}[[MPR]], i{{64|32}} 0, i{{64|32}} 1
+// CK3-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [10 x %class.C]**
+// CK3-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to %class.C**
+// CK3-DAG: store [10 x %class.C]* [[CVAL]], [10 x %class.C]** [[CBP2]]
+// CK3-DAG: [[CVALGEP:%.+]] = getelementptr inbounds {{.+}}[[CVAL]], i{{64|32}} 0, i{{64|32}} 0
+// CK3-DAG: store %class.C* [[CVALGEP]], %class.C** [[CP2]]
+// CK3-DAG: store i8* bitcast (void (i8*, i8*, i8*, i64, i64, i8*)* [[MPRFUNC]] to i8*), i8** [[MPR2]]
+// CK3: call void [[KERNEL:@.+]](%class.B* [[BVAL]], [10 x %class.C]* [[CVAL]])
+#pragma omp target map(mapper(id), tofrom \
+                       : c [0:10], b.c)
   for (int i = 0; i < 10; i++) {
     b.c.a += ++c[i].a;
   }

diff --git a/clang/test/OpenMP/declare_target_link_codegen.cpp b/clang/test/OpenMP/declare_target_link_codegen.cpp
index ca4ba07..3e6ca6a 100644 (file)
--- a/clang/test/OpenMP/declare_target_link_codegen.cpp
+++ b/clang/test/OpenMP/declare_target_link_codegen.cpp
@@ -77,9 +77,9 @@ int maini1() {
 
 // HOST: [[BP0:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[BASEPTRS]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
 // HOST: [[P0:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[PTRS]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
-// HOST: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 3, i8** [[BP0]], i8** [[P0]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* [[SIZES]], i{{[0-9]+}} 0, i{{[0-9]+}} 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* [[MAPTYPES]], i{{[0-9]+}} 0, i{{[0-9]+}} 0), i8** null, i8** null)
+// HOST: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 // HOST: call void @__omp_offloading_{{.*}}_{{.*}}_{{.*}}maini1{{.*}}_l42(i32* %{{[^,]+}})
-// HOST: call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @.__omp_offloading_{{.+}}_l47.region_id, i32 2, {{.+}})
+// HOST: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 0, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 
 // HOST: define internal void @__omp_offloading_{{.*}}_{{.*}}maini1{{.*}}_l42(i32* noundef nonnull align {{[0-9]+}} dereferenceable{{.*}})
 // HOST: [[C:%.*]] = load i32, i32* @c,

diff --git a/clang/test/OpenMP/distribute_codegen.cpp b/clang/test/OpenMP/distribute_codegen.cpp
index 855efca..ab3a576 100644 (file)
--- a/clang/test/OpenMP/distribute_codegen.cpp
+++ b/clang/test/OpenMP/distribute_codegen.cpp
@@ -170,9 +170,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 4571424)
-// CHECK1-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l56.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK1-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK1-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l56(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -343,9 +360,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 4571424)
-// CHECK1-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l68.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK1-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK1-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l68(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -516,9 +550,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 16908289)
-// CHECK1-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l80.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK1-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK1-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l80(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -692,9 +743,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK1-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP11]], 1
 // CHECK1-NEXT:    [[TMP12:%.*]] = zext i32 [[ADD5]] to i64
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP12]])
-// CHECK1-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l92.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK1-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP13]], align 4
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP14]], align 4
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP22:%.*]] = icmp ne i32 [[TMP21]], 0
+// CHECK1-NEXT:    br i1 [[TMP22]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l92(i64 [[TMP1]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -837,9 +905,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l108.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l108(i64 [[TMP1]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1000,9 +1085,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 4571424)
-// CHECK3-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l56.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK3-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK3-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l56(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1169,9 +1271,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 4571424)
-// CHECK3-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l68.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK3-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK3-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l68(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1338,9 +1457,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 16908289)
-// CHECK3-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l80.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK3-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK3-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l80(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1510,9 +1646,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK3-NEXT:    [[ADD5:%.*]] = add nsw i32 [[TMP11]], 1
 // CHECK3-NEXT:    [[TMP12:%.*]] = zext i32 [[ADD5]] to i64
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP12]])
-// CHECK3-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l92.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK3-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP19]], align 4
+// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP22:%.*]] = icmp ne i32 [[TMP21]], 0
+// CHECK3-NEXT:    br i1 [[TMP22]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l92(i32 [[TMP1]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1655,9 +1808,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l108.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK3-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l108(i32 [[TMP1]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/distribute_firstprivate_codegen.cpp b/clang/test/OpenMP/distribute_firstprivate_codegen.cpp
index 4abfc4f..79e1e0e 100644 (file)
--- a/clang/test/OpenMP/distribute_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/distribute_firstprivate_codegen.cpp
@@ -573,9 +573,26 @@ int main() {
 // CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK9-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK9-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -583,18 +600,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done3:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP39]]
+// CHECK9-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -864,27 +881,44 @@ int main() {
 // CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK9-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK9-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP32]]
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1235,9 +1269,26 @@ int main() {
 // CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK11-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK11-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1245,18 +1296,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP39]]
+// CHECK11-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1521,27 +1572,44 @@ int main() {
 // CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK11-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK11-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP32]]
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/distribute_lastprivate_codegen.cpp b/clang/test/OpenMP/distribute_lastprivate_codegen.cpp
index 7420e1b..ffa4b42 100644 (file)
--- a/clang/test/OpenMP/distribute_lastprivate_codegen.cpp
+++ b/clang/test/OpenMP/distribute_lastprivate_codegen.cpp
@@ -558,9 +558,26 @@ int main() {
 // CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK9-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK9-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -568,18 +585,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done3:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP39]]
+// CHECK9-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -867,27 +884,44 @@ int main() {
 // CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK9-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK9-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP32]]
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1257,9 +1291,26 @@ int main() {
 // CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK11-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK11-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1267,18 +1318,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP39]]
+// CHECK11-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1561,27 +1612,44 @@ int main() {
 // CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK11-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK11-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP32]]
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/distribute_parallel_for_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_codegen.cpp
index 5402da8..85612f4 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_codegen.cpp
@@ -4310,43 +4310,43 @@ int main() {
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[CH_CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED18:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED32:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS34:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS35:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS36:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP37:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_38:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_39:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[CH_CASTED46:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED48:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS50:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS51:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS52:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP53:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_54:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_55:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED62:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS64:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS65:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS66:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP67:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_68:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_69:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[CH_CASTED76:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED78:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS80:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS81:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS82:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP83:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_84:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_85:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED34:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS36:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS37:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS38:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP39:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_40:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_41:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[CH_CASTED49:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[N_CASTED51:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS53:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS54:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS55:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP56:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_57:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_58:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED66:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS68:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS69:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS70:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP71:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_72:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_73:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[CH_CASTED81:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[N_CASTED83:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS85:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS86:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS87:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP88:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_89:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_90:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    store i32 10000, i32* [[N]], align 4
 // CHECK9-NEXT:    store i32 100, i32* [[CH]], align 4
@@ -4402,403 +4402,522 @@ int main() {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP29]], 1
 // CHECK9-NEXT:    [[TMP30:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP30]])
-// CHECK9-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l369.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK9-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP32]], align 4
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP25]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP26]], i8*** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l369.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK9-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l369(i64 [[TMP1]], double* [[TMP2]], double* [[TMP3]], double* [[TMP4]]) #[[ATTR2:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N]], align 4
 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP33]], i32* [[CONV4]], align 4
-// CHECK9-NEXT:    [[TMP34:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK9-NEXT:    [[TMP35:%.*]] = load double*, double** [[A]], align 8
-// CHECK9-NEXT:    [[TMP36:%.*]] = load double*, double** [[B]], align 8
-// CHECK9-NEXT:    [[TMP37:%.*]] = load double*, double** [[C]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP41]], align 8
-// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP42]], align 8
-// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to double**
-// CHECK9-NEXT:    store double* [[TMP35]], double** [[TMP44]], align 8
-// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to double**
-// CHECK9-NEXT:    store double* [[TMP35]], double** [[TMP46]], align 8
-// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP47]], align 8
-// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to double**
-// CHECK9-NEXT:    store double* [[TMP36]], double** [[TMP49]], align 8
-// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to double**
-// CHECK9-NEXT:    store double* [[TMP36]], double** [[TMP51]], align 8
-// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP52]], align 8
-// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 3
+// CHECK9-NEXT:    store i32 [[TMP41]], i32* [[CONV4]], align 4
+// CHECK9-NEXT:    [[TMP42:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = load double*, double** [[A]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = load double*, double** [[B]], align 8
+// CHECK9-NEXT:    [[TMP45:%.*]] = load double*, double** [[C]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP49]], align 8
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to double**
+// CHECK9-NEXT:    store double* [[TMP43]], double** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
 // CHECK9-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to double**
-// CHECK9-NEXT:    store double* [[TMP37]], double** [[TMP54]], align 8
-// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to double**
-// CHECK9-NEXT:    store double* [[TMP37]], double** [[TMP56]], align 8
-// CHECK9-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP57]], align 8
-// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP60:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP60]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK9-NEXT:    [[TMP61:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK9-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP61]], 0
+// CHECK9-NEXT:    store double* [[TMP43]], double** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to double**
+// CHECK9-NEXT:    store double* [[TMP44]], double** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to double**
+// CHECK9-NEXT:    store double* [[TMP44]], double** [[TMP59]], align 8
+// CHECK9-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP60]], align 8
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to double**
+// CHECK9-NEXT:    store double* [[TMP45]], double** [[TMP62]], align 8
+// CHECK9-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP64:%.*]] = bitcast i8** [[TMP63]] to double**
+// CHECK9-NEXT:    store double* [[TMP45]], double** [[TMP64]], align 8
+// CHECK9-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP65]], align 8
+// CHECK9-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP68]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK9-NEXT:    [[TMP69:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK9-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP69]], 0
 // CHECK9-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK9-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK9-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[TMP62:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP62]], 1
-// CHECK9-NEXT:    [[TMP63:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP63]])
-// CHECK9-NEXT:    [[TMP64:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l408.region_id, i32 4, i8** [[TMP58]], i8** [[TMP59]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP65:%.*]] = icmp ne i32 [[TMP64]], 0
-// CHECK9-NEXT:    br i1 [[TMP65]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK9:       omp_offload.failed15:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l408(i64 [[TMP34]], double* [[TMP35]], double* [[TMP36]], double* [[TMP37]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK9:       omp_offload.cont16:
-// CHECK9-NEXT:    [[TMP66:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK9-NEXT:    [[CONV17:%.*]] = bitcast i64* [[CH_CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP66]], i32* [[CONV17]], align 4
-// CHECK9-NEXT:    [[TMP67:%.*]] = load i64, i64* [[CH_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV19:%.*]] = bitcast i64* [[N_CASTED18]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP68]], i32* [[CONV19]], align 4
-// CHECK9-NEXT:    [[TMP69:%.*]] = load i64, i64* [[N_CASTED18]], align 8
-// CHECK9-NEXT:    [[TMP70:%.*]] = load double*, double** [[A]], align 8
-// CHECK9-NEXT:    [[TMP71:%.*]] = load double*, double** [[B]], align 8
-// CHECK9-NEXT:    [[TMP72:%.*]] = load double*, double** [[C]], align 8
-// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP67]], i64* [[TMP74]], align 8
-// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP67]], i64* [[TMP76]], align 8
-// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP77]], align 8
-// CHECK9-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP69]], i64* [[TMP79]], align 8
-// CHECK9-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP69]], i64* [[TMP81]], align 8
-// CHECK9-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP82]], align 8
-// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to double**
-// CHECK9-NEXT:    store double* [[TMP70]], double** [[TMP84]], align 8
-// CHECK9-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to double**
-// CHECK9-NEXT:    store double* [[TMP70]], double** [[TMP86]], align 8
-// CHECK9-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP87]], align 8
-// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to double**
-// CHECK9-NEXT:    store double* [[TMP71]], double** [[TMP89]], align 8
-// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to double**
-// CHECK9-NEXT:    store double* [[TMP71]], double** [[TMP91]], align 8
-// CHECK9-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP92]], align 8
-// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to double**
-// CHECK9-NEXT:    store double* [[TMP72]], double** [[TMP94]], align 8
-// CHECK9-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to double**
-// CHECK9-NEXT:    store double* [[TMP72]], double** [[TMP96]], align 8
-// CHECK9-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 4
-// CHECK9-NEXT:    store i8* null, i8** [[TMP97]], align 8
-// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP100]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK9-NEXT:    [[TMP101:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK9-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP101]], 0
-// CHECK9-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
-// CHECK9-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
-// CHECK9-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK9-NEXT:    [[TMP102:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK9-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP102]], 1
-// CHECK9-NEXT:    [[TMP103:%.*]] = zext i32 [[ADD29]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP103]])
-// CHECK9-NEXT:    [[TMP104:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l447.region_id, i32 5, i8** [[TMP98]], i8** [[TMP99]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP105:%.*]] = icmp ne i32 [[TMP104]], 0
-// CHECK9-NEXT:    br i1 [[TMP105]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK9:       omp_offload.failed30:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l447(i64 [[TMP67]], i64 [[TMP69]], double* [[TMP70]], double* [[TMP71]], double* [[TMP72]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK9:       omp_offload.cont31:
-// CHECK9-NEXT:    [[TMP106:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV33:%.*]] = bitcast i64* [[N_CASTED32]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP106]], i32* [[CONV33]], align 4
-// CHECK9-NEXT:    [[TMP107:%.*]] = load i64, i64* [[N_CASTED32]], align 8
-// CHECK9-NEXT:    [[TMP108:%.*]] = load double*, double** [[A]], align 8
-// CHECK9-NEXT:    [[TMP109:%.*]] = load double*, double** [[B]], align 8
-// CHECK9-NEXT:    [[TMP110:%.*]] = load double*, double** [[C]], align 8
-// CHECK9-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP107]], i64* [[TMP112]], align 8
-// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP107]], i64* [[TMP114]], align 8
-// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP115]], align 8
-// CHECK9-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to double**
-// CHECK9-NEXT:    store double* [[TMP108]], double** [[TMP117]], align 8
-// CHECK9-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP119:%.*]] = bitcast i8** [[TMP118]] to double**
-// CHECK9-NEXT:    store double* [[TMP108]], double** [[TMP119]], align 8
-// CHECK9-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP120]], align 8
-// CHECK9-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to double**
-// CHECK9-NEXT:    store double* [[TMP109]], double** [[TMP122]], align 8
-// CHECK9-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to double**
-// CHECK9-NEXT:    store double* [[TMP109]], double** [[TMP124]], align 8
-// CHECK9-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP125]], align 8
-// CHECK9-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to double**
-// CHECK9-NEXT:    store double* [[TMP110]], double** [[TMP127]], align 8
-// CHECK9-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to double**
-// CHECK9-NEXT:    store double* [[TMP110]], double** [[TMP129]], align 8
-// CHECK9-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP130]], align 8
-// CHECK9-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP133:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP133]], i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK9-NEXT:    [[TMP134:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK9-NEXT:    [[SUB40:%.*]] = sub nsw i32 [[TMP134]], 0
-// CHECK9-NEXT:    [[DIV41:%.*]] = sdiv i32 [[SUB40]], 1
-// CHECK9-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[DIV41]], 1
-// CHECK9-NEXT:    store i32 [[SUB42]], i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK9-NEXT:    [[TMP135:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK9-NEXT:    [[ADD43:%.*]] = add nsw i32 [[TMP135]], 1
-// CHECK9-NEXT:    [[TMP136:%.*]] = zext i32 [[ADD43]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP136]])
-// CHECK9-NEXT:    [[TMP137:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l478.region_id, i32 4, i8** [[TMP131]], i8** [[TMP132]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP138:%.*]] = icmp ne i32 [[TMP137]], 0
-// CHECK9-NEXT:    br i1 [[TMP138]], label [[OMP_OFFLOAD_FAILED44:%.*]], label [[OMP_OFFLOAD_CONT45:%.*]]
-// CHECK9:       omp_offload.failed44:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l478(i64 [[TMP107]], double* [[TMP108]], double* [[TMP109]], double* [[TMP110]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT45]]
-// CHECK9:       omp_offload.cont45:
-// CHECK9-NEXT:    [[TMP139:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK9-NEXT:    [[CONV47:%.*]] = bitcast i64* [[CH_CASTED46]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP139]], i32* [[CONV47]], align 4
-// CHECK9-NEXT:    [[TMP140:%.*]] = load i64, i64* [[CH_CASTED46]], align 8
-// CHECK9-NEXT:    [[TMP141:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV49:%.*]] = bitcast i64* [[N_CASTED48]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP141]], i32* [[CONV49]], align 4
-// CHECK9-NEXT:    [[TMP142:%.*]] = load i64, i64* [[N_CASTED48]], align 8
-// CHECK9-NEXT:    [[TMP143:%.*]] = load double*, double** [[A]], align 8
-// CHECK9-NEXT:    [[TMP144:%.*]] = load double*, double** [[B]], align 8
-// CHECK9-NEXT:    [[TMP145:%.*]] = load double*, double** [[C]], align 8
-// CHECK9-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP140]], i64* [[TMP147]], align 8
-// CHECK9-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP140]], i64* [[TMP149]], align 8
-// CHECK9-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP150]], align 8
-// CHECK9-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP142]], i64* [[TMP152]], align 8
-// CHECK9-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP142]], i64* [[TMP154]], align 8
-// CHECK9-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP155]], align 8
-// CHECK9-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to double**
-// CHECK9-NEXT:    store double* [[TMP143]], double** [[TMP157]], align 8
-// CHECK9-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to double**
-// CHECK9-NEXT:    store double* [[TMP143]], double** [[TMP159]], align 8
-// CHECK9-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP160]], align 8
-// CHECK9-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP162:%.*]] = bitcast i8** [[TMP161]] to double**
-// CHECK9-NEXT:    store double* [[TMP144]], double** [[TMP162]], align 8
-// CHECK9-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP164:%.*]] = bitcast i8** [[TMP163]] to double**
-// CHECK9-NEXT:    store double* [[TMP144]], double** [[TMP164]], align 8
-// CHECK9-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP165]], align 8
-// CHECK9-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to double**
-// CHECK9-NEXT:    store double* [[TMP145]], double** [[TMP167]], align 8
-// CHECK9-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to double**
-// CHECK9-NEXT:    store double* [[TMP145]], double** [[TMP169]], align 8
-// CHECK9-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 4
-// CHECK9-NEXT:    store i8* null, i8** [[TMP170]], align 8
-// CHECK9-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP70:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP70]], 1
+// CHECK9-NEXT:    [[TMP71:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP71]])
+// CHECK9-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP72]], align 4
+// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP73]], align 4
+// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP66]], i8*** [[TMP74]], align 8
+// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP67]], i8*** [[TMP75]], align 8
+// CHECK9-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP76]], align 8
+// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP77]], align 8
+// CHECK9-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP78]], align 8
+// CHECK9-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP79]], align 8
+// CHECK9-NEXT:    [[TMP80:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l408.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK9-NEXT:    [[TMP81:%.*]] = icmp ne i32 [[TMP80]], 0
+// CHECK9-NEXT:    br i1 [[TMP81]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK9:       omp_offload.failed16:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l408(i64 [[TMP42]], double* [[TMP43]], double* [[TMP44]], double* [[TMP45]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK9:       omp_offload.cont17:
+// CHECK9-NEXT:    [[TMP82:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK9-NEXT:    [[CONV18:%.*]] = bitcast i64* [[CH_CASTED]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP82]], i32* [[CONV18]], align 4
+// CHECK9-NEXT:    [[TMP83:%.*]] = load i64, i64* [[CH_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP84:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP84]], i32* [[CONV20]], align 4
+// CHECK9-NEXT:    [[TMP85:%.*]] = load i64, i64* [[N_CASTED19]], align 8
+// CHECK9-NEXT:    [[TMP86:%.*]] = load double*, double** [[A]], align 8
+// CHECK9-NEXT:    [[TMP87:%.*]] = load double*, double** [[B]], align 8
+// CHECK9-NEXT:    [[TMP88:%.*]] = load double*, double** [[C]], align 8
+// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP83]], i64* [[TMP90]], align 8
+// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP83]], i64* [[TMP92]], align 8
+// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP93]], align 8
+// CHECK9-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP85]], i64* [[TMP95]], align 8
+// CHECK9-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP85]], i64* [[TMP97]], align 8
+// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP98]], align 8
+// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP100:%.*]] = bitcast i8** [[TMP99]] to double**
+// CHECK9-NEXT:    store double* [[TMP86]], double** [[TMP100]], align 8
+// CHECK9-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to double**
+// CHECK9-NEXT:    store double* [[TMP86]], double** [[TMP102]], align 8
+// CHECK9-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP103]], align 8
+// CHECK9-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to double**
+// CHECK9-NEXT:    store double* [[TMP87]], double** [[TMP105]], align 8
+// CHECK9-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to double**
+// CHECK9-NEXT:    store double* [[TMP87]], double** [[TMP107]], align 8
+// CHECK9-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP108]], align 8
+// CHECK9-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to double**
+// CHECK9-NEXT:    store double* [[TMP88]], double** [[TMP110]], align 8
+// CHECK9-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to double**
+// CHECK9-NEXT:    store double* [[TMP88]], double** [[TMP112]], align 8
+// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 4
+// CHECK9-NEXT:    store i8* null, i8** [[TMP113]], align 8
+// CHECK9-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP116:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP116]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK9-NEXT:    [[TMP117:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK9-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP117]], 0
+// CHECK9-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
+// CHECK9-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
+// CHECK9-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK9-NEXT:    [[TMP118:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK9-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP118]], 1
+// CHECK9-NEXT:    [[TMP119:%.*]] = zext i32 [[ADD30]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP119]])
+// CHECK9-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP120]], align 4
+// CHECK9-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP121]], align 4
+// CHECK9-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP114]], i8*** [[TMP122]], align 8
+// CHECK9-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP115]], i8*** [[TMP123]], align 8
+// CHECK9-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP124]], align 8
+// CHECK9-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP125]], align 8
+// CHECK9-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP126]], align 8
+// CHECK9-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP127]], align 8
+// CHECK9-NEXT:    [[TMP128:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l447.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK9-NEXT:    [[TMP129:%.*]] = icmp ne i32 [[TMP128]], 0
+// CHECK9-NEXT:    br i1 [[TMP129]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK9:       omp_offload.failed32:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l447(i64 [[TMP83]], i64 [[TMP85]], double* [[TMP86]], double* [[TMP87]], double* [[TMP88]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK9:       omp_offload.cont33:
+// CHECK9-NEXT:    [[TMP130:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV35:%.*]] = bitcast i64* [[N_CASTED34]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP130]], i32* [[CONV35]], align 4
+// CHECK9-NEXT:    [[TMP131:%.*]] = load i64, i64* [[N_CASTED34]], align 8
+// CHECK9-NEXT:    [[TMP132:%.*]] = load double*, double** [[A]], align 8
+// CHECK9-NEXT:    [[TMP133:%.*]] = load double*, double** [[B]], align 8
+// CHECK9-NEXT:    [[TMP134:%.*]] = load double*, double** [[C]], align 8
+// CHECK9-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP131]], i64* [[TMP136]], align 8
+// CHECK9-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP131]], i64* [[TMP138]], align 8
+// CHECK9-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP139]], align 8
+// CHECK9-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to double**
+// CHECK9-NEXT:    store double* [[TMP132]], double** [[TMP141]], align 8
+// CHECK9-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to double**
+// CHECK9-NEXT:    store double* [[TMP132]], double** [[TMP143]], align 8
+// CHECK9-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP144]], align 8
+// CHECK9-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to double**
+// CHECK9-NEXT:    store double* [[TMP133]], double** [[TMP146]], align 8
+// CHECK9-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to double**
+// CHECK9-NEXT:    store double* [[TMP133]], double** [[TMP148]], align 8
+// CHECK9-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP149]], align 8
+// CHECK9-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to double**
+// CHECK9-NEXT:    store double* [[TMP134]], double** [[TMP151]], align 8
+// CHECK9-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to double**
+// CHECK9-NEXT:    store double* [[TMP134]], double** [[TMP153]], align 8
+// CHECK9-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP154]], align 8
+// CHECK9-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP157:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP157]], i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK9-NEXT:    [[TMP158:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK9-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[TMP158]], 0
+// CHECK9-NEXT:    [[DIV43:%.*]] = sdiv i32 [[SUB42]], 1
+// CHECK9-NEXT:    [[SUB44:%.*]] = sub nsw i32 [[DIV43]], 1
+// CHECK9-NEXT:    store i32 [[SUB44]], i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK9-NEXT:    [[TMP159:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK9-NEXT:    [[ADD45:%.*]] = add nsw i32 [[TMP159]], 1
+// CHECK9-NEXT:    [[TMP160:%.*]] = zext i32 [[ADD45]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP160]])
+// CHECK9-NEXT:    [[KERNEL_ARGS46:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP161]], align 4
+// CHECK9-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP162]], align 4
+// CHECK9-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP155]], i8*** [[TMP163]], align 8
+// CHECK9-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP156]], i8*** [[TMP164]], align 8
+// CHECK9-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP165]], align 8
+// CHECK9-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP166]], align 8
+// CHECK9-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP167]], align 8
+// CHECK9-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP168]], align 8
+// CHECK9-NEXT:    [[TMP169:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l478.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]])
+// CHECK9-NEXT:    [[TMP170:%.*]] = icmp ne i32 [[TMP169]], 0
+// CHECK9-NEXT:    br i1 [[TMP170]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
+// CHECK9:       omp_offload.failed47:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l478(i64 [[TMP131]], double* [[TMP132]], double* [[TMP133]], double* [[TMP134]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
+// CHECK9:       omp_offload.cont48:
+// CHECK9-NEXT:    [[TMP171:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK9-NEXT:    [[CONV50:%.*]] = bitcast i64* [[CH_CASTED49]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP171]], i32* [[CONV50]], align 4
+// CHECK9-NEXT:    [[TMP172:%.*]] = load i64, i64* [[CH_CASTED49]], align 8
 // CHECK9-NEXT:    [[TMP173:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP173]], i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK9-NEXT:    [[TMP174:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK9-NEXT:    [[SUB56:%.*]] = sub nsw i32 [[TMP174]], 0
-// CHECK9-NEXT:    [[DIV57:%.*]] = sdiv i32 [[SUB56]], 1
-// CHECK9-NEXT:    [[SUB58:%.*]] = sub nsw i32 [[DIV57]], 1
-// CHECK9-NEXT:    store i32 [[SUB58]], i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK9-NEXT:    [[TMP175:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK9-NEXT:    [[ADD59:%.*]] = add nsw i32 [[TMP175]], 1
-// CHECK9-NEXT:    [[TMP176:%.*]] = zext i32 [[ADD59]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP176]])
-// CHECK9-NEXT:    [[TMP177:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l506.region_id, i32 5, i8** [[TMP171]], i8** [[TMP172]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP178:%.*]] = icmp ne i32 [[TMP177]], 0
-// CHECK9-NEXT:    br i1 [[TMP178]], label [[OMP_OFFLOAD_FAILED60:%.*]], label [[OMP_OFFLOAD_CONT61:%.*]]
-// CHECK9:       omp_offload.failed60:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l506(i64 [[TMP140]], i64 [[TMP142]], double* [[TMP143]], double* [[TMP144]], double* [[TMP145]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT61]]
-// CHECK9:       omp_offload.cont61:
-// CHECK9-NEXT:    [[TMP179:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV63:%.*]] = bitcast i64* [[N_CASTED62]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP179]], i32* [[CONV63]], align 4
-// CHECK9-NEXT:    [[TMP180:%.*]] = load i64, i64* [[N_CASTED62]], align 8
-// CHECK9-NEXT:    [[TMP181:%.*]] = load double*, double** [[A]], align 8
-// CHECK9-NEXT:    [[TMP182:%.*]] = load double*, double** [[B]], align 8
-// CHECK9-NEXT:    [[TMP183:%.*]] = load double*, double** [[C]], align 8
-// CHECK9-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP185:%.*]] = bitcast i8** [[TMP184]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP180]], i64* [[TMP185]], align 8
-// CHECK9-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP187:%.*]] = bitcast i8** [[TMP186]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP180]], i64* [[TMP187]], align 8
-// CHECK9-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP188]], align 8
-// CHECK9-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP190:%.*]] = bitcast i8** [[TMP189]] to double**
-// CHECK9-NEXT:    store double* [[TMP181]], double** [[TMP190]], align 8
-// CHECK9-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP192:%.*]] = bitcast i8** [[TMP191]] to double**
-// CHECK9-NEXT:    store double* [[TMP181]], double** [[TMP192]], align 8
-// CHECK9-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP193]], align 8
-// CHECK9-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP195:%.*]] = bitcast i8** [[TMP194]] to double**
-// CHECK9-NEXT:    store double* [[TMP182]], double** [[TMP195]], align 8
-// CHECK9-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP197:%.*]] = bitcast i8** [[TMP196]] to double**
-// CHECK9-NEXT:    store double* [[TMP182]], double** [[TMP197]], align 8
-// CHECK9-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP198]], align 8
-// CHECK9-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP200:%.*]] = bitcast i8** [[TMP199]] to double**
-// CHECK9-NEXT:    store double* [[TMP183]], double** [[TMP200]], align 8
-// CHECK9-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP202:%.*]] = bitcast i8** [[TMP201]] to double**
-// CHECK9-NEXT:    store double* [[TMP183]], double** [[TMP202]], align 8
-// CHECK9-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP203]], align 8
-// CHECK9-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP206:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP206]], i32* [[DOTCAPTURE_EXPR_68]], align 4
-// CHECK9-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_68]], align 4
-// CHECK9-NEXT:    [[SUB70:%.*]] = sub nsw i32 [[TMP207]], 0
-// CHECK9-NEXT:    [[DIV71:%.*]] = sdiv i32 [[SUB70]], 1
-// CHECK9-NEXT:    [[SUB72:%.*]] = sub nsw i32 [[DIV71]], 1
-// CHECK9-NEXT:    store i32 [[SUB72]], i32* [[DOTCAPTURE_EXPR_69]], align 4
-// CHECK9-NEXT:    [[TMP208:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_69]], align 4
-// CHECK9-NEXT:    [[ADD73:%.*]] = add nsw i32 [[TMP208]], 1
-// CHECK9-NEXT:    [[TMP209:%.*]] = zext i32 [[ADD73]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP209]])
-// CHECK9-NEXT:    [[TMP210:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l536.region_id, i32 4, i8** [[TMP204]], i8** [[TMP205]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP211:%.*]] = icmp ne i32 [[TMP210]], 0
-// CHECK9-NEXT:    br i1 [[TMP211]], label [[OMP_OFFLOAD_FAILED74:%.*]], label [[OMP_OFFLOAD_CONT75:%.*]]
-// CHECK9:       omp_offload.failed74:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l536(i64 [[TMP180]], double* [[TMP181]], double* [[TMP182]], double* [[TMP183]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT75]]
-// CHECK9:       omp_offload.cont75:
-// CHECK9-NEXT:    [[TMP212:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK9-NEXT:    [[CONV77:%.*]] = bitcast i64* [[CH_CASTED76]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP212]], i32* [[CONV77]], align 4
-// CHECK9-NEXT:    [[TMP213:%.*]] = load i64, i64* [[CH_CASTED76]], align 8
-// CHECK9-NEXT:    [[TMP214:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV79:%.*]] = bitcast i64* [[N_CASTED78]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP214]], i32* [[CONV79]], align 4
-// CHECK9-NEXT:    [[TMP215:%.*]] = load i64, i64* [[N_CASTED78]], align 8
-// CHECK9-NEXT:    [[TMP216:%.*]] = load double*, double** [[A]], align 8
-// CHECK9-NEXT:    [[TMP217:%.*]] = load double*, double** [[B]], align 8
-// CHECK9-NEXT:    [[TMP218:%.*]] = load double*, double** [[C]], align 8
-// CHECK9-NEXT:    [[TMP219:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP220:%.*]] = bitcast i8** [[TMP219]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP213]], i64* [[TMP220]], align 8
-// CHECK9-NEXT:    [[TMP221:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP222:%.*]] = bitcast i8** [[TMP221]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP213]], i64* [[TMP222]], align 8
-// CHECK9-NEXT:    [[TMP223:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP223]], align 8
-// CHECK9-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 1
+// CHECK9-NEXT:    [[CONV52:%.*]] = bitcast i64* [[N_CASTED51]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP173]], i32* [[CONV52]], align 4
+// CHECK9-NEXT:    [[TMP174:%.*]] = load i64, i64* [[N_CASTED51]], align 8
+// CHECK9-NEXT:    [[TMP175:%.*]] = load double*, double** [[A]], align 8
+// CHECK9-NEXT:    [[TMP176:%.*]] = load double*, double** [[B]], align 8
+// CHECK9-NEXT:    [[TMP177:%.*]] = load double*, double** [[C]], align 8
+// CHECK9-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP172]], i64* [[TMP179]], align 8
+// CHECK9-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP172]], i64* [[TMP181]], align 8
+// CHECK9-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP182]], align 8
+// CHECK9-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP184:%.*]] = bitcast i8** [[TMP183]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP174]], i64* [[TMP184]], align 8
+// CHECK9-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP186:%.*]] = bitcast i8** [[TMP185]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP174]], i64* [[TMP186]], align 8
+// CHECK9-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP187]], align 8
+// CHECK9-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP189:%.*]] = bitcast i8** [[TMP188]] to double**
+// CHECK9-NEXT:    store double* [[TMP175]], double** [[TMP189]], align 8
+// CHECK9-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP191:%.*]] = bitcast i8** [[TMP190]] to double**
+// CHECK9-NEXT:    store double* [[TMP175]], double** [[TMP191]], align 8
+// CHECK9-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP192]], align 8
+// CHECK9-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP194:%.*]] = bitcast i8** [[TMP193]] to double**
+// CHECK9-NEXT:    store double* [[TMP176]], double** [[TMP194]], align 8
+// CHECK9-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP196:%.*]] = bitcast i8** [[TMP195]] to double**
+// CHECK9-NEXT:    store double* [[TMP176]], double** [[TMP196]], align 8
+// CHECK9-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP197]], align 8
+// CHECK9-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP199:%.*]] = bitcast i8** [[TMP198]] to double**
+// CHECK9-NEXT:    store double* [[TMP177]], double** [[TMP199]], align 8
+// CHECK9-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP201:%.*]] = bitcast i8** [[TMP200]] to double**
+// CHECK9-NEXT:    store double* [[TMP177]], double** [[TMP201]], align 8
+// CHECK9-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 4
+// CHECK9-NEXT:    store i8* null, i8** [[TMP202]], align 8
+// CHECK9-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP205:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP205]], i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK9-NEXT:    [[TMP206:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK9-NEXT:    [[SUB59:%.*]] = sub nsw i32 [[TMP206]], 0
+// CHECK9-NEXT:    [[DIV60:%.*]] = sdiv i32 [[SUB59]], 1
+// CHECK9-NEXT:    [[SUB61:%.*]] = sub nsw i32 [[DIV60]], 1
+// CHECK9-NEXT:    store i32 [[SUB61]], i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK9-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK9-NEXT:    [[ADD62:%.*]] = add nsw i32 [[TMP207]], 1
+// CHECK9-NEXT:    [[TMP208:%.*]] = zext i32 [[ADD62]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP208]])
+// CHECK9-NEXT:    [[KERNEL_ARGS63:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP209:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP209]], align 4
+// CHECK9-NEXT:    [[TMP210:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP210]], align 4
+// CHECK9-NEXT:    [[TMP211:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP203]], i8*** [[TMP211]], align 8
+// CHECK9-NEXT:    [[TMP212:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP204]], i8*** [[TMP212]], align 8
+// CHECK9-NEXT:    [[TMP213:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP213]], align 8
+// CHECK9-NEXT:    [[TMP214:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP214]], align 8
+// CHECK9-NEXT:    [[TMP215:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP215]], align 8
+// CHECK9-NEXT:    [[TMP216:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP216]], align 8
+// CHECK9-NEXT:    [[TMP217:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l506.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]])
+// CHECK9-NEXT:    [[TMP218:%.*]] = icmp ne i32 [[TMP217]], 0
+// CHECK9-NEXT:    br i1 [[TMP218]], label [[OMP_OFFLOAD_FAILED64:%.*]], label [[OMP_OFFLOAD_CONT65:%.*]]
+// CHECK9:       omp_offload.failed64:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l506(i64 [[TMP172]], i64 [[TMP174]], double* [[TMP175]], double* [[TMP176]], double* [[TMP177]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT65]]
+// CHECK9:       omp_offload.cont65:
+// CHECK9-NEXT:    [[TMP219:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV67:%.*]] = bitcast i64* [[N_CASTED66]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP219]], i32* [[CONV67]], align 4
+// CHECK9-NEXT:    [[TMP220:%.*]] = load i64, i64* [[N_CASTED66]], align 8
+// CHECK9-NEXT:    [[TMP221:%.*]] = load double*, double** [[A]], align 8
+// CHECK9-NEXT:    [[TMP222:%.*]] = load double*, double** [[B]], align 8
+// CHECK9-NEXT:    [[TMP223:%.*]] = load double*, double** [[C]], align 8
+// CHECK9-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP225:%.*]] = bitcast i8** [[TMP224]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP215]], i64* [[TMP225]], align 8
-// CHECK9-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 1
+// CHECK9-NEXT:    store i64 [[TMP220]], i64* [[TMP225]], align 8
+// CHECK9-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP227:%.*]] = bitcast i8** [[TMP226]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP215]], i64* [[TMP227]], align 8
-// CHECK9-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 1
+// CHECK9-NEXT:    store i64 [[TMP220]], i64* [[TMP227]], align 8
+// CHECK9-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 0
 // CHECK9-NEXT:    store i8* null, i8** [[TMP228]], align 8
-// CHECK9-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 1
 // CHECK9-NEXT:    [[TMP230:%.*]] = bitcast i8** [[TMP229]] to double**
-// CHECK9-NEXT:    store double* [[TMP216]], double** [[TMP230]], align 8
-// CHECK9-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 2
+// CHECK9-NEXT:    store double* [[TMP221]], double** [[TMP230]], align 8
+// CHECK9-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 1
 // CHECK9-NEXT:    [[TMP232:%.*]] = bitcast i8** [[TMP231]] to double**
-// CHECK9-NEXT:    store double* [[TMP216]], double** [[TMP232]], align 8
-// CHECK9-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 2
+// CHECK9-NEXT:    store double* [[TMP221]], double** [[TMP232]], align 8
+// CHECK9-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 1
 // CHECK9-NEXT:    store i8* null, i8** [[TMP233]], align 8
-// CHECK9-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 2
 // CHECK9-NEXT:    [[TMP235:%.*]] = bitcast i8** [[TMP234]] to double**
-// CHECK9-NEXT:    store double* [[TMP217]], double** [[TMP235]], align 8
-// CHECK9-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 3
+// CHECK9-NEXT:    store double* [[TMP222]], double** [[TMP235]], align 8
+// CHECK9-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 2
 // CHECK9-NEXT:    [[TMP237:%.*]] = bitcast i8** [[TMP236]] to double**
-// CHECK9-NEXT:    store double* [[TMP217]], double** [[TMP237]], align 8
-// CHECK9-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 3
+// CHECK9-NEXT:    store double* [[TMP222]], double** [[TMP237]], align 8
+// CHECK9-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 2
 // CHECK9-NEXT:    store i8* null, i8** [[TMP238]], align 8
-// CHECK9-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 3
 // CHECK9-NEXT:    [[TMP240:%.*]] = bitcast i8** [[TMP239]] to double**
-// CHECK9-NEXT:    store double* [[TMP218]], double** [[TMP240]], align 8
-// CHECK9-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 4
+// CHECK9-NEXT:    store double* [[TMP223]], double** [[TMP240]], align 8
+// CHECK9-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 3
 // CHECK9-NEXT:    [[TMP242:%.*]] = bitcast i8** [[TMP241]] to double**
-// CHECK9-NEXT:    store double* [[TMP218]], double** [[TMP242]], align 8
-// CHECK9-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 4
+// CHECK9-NEXT:    store double* [[TMP223]], double** [[TMP242]], align 8
+// CHECK9-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 3
 // CHECK9-NEXT:    store i8* null, i8** [[TMP243]], align 8
-// CHECK9-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP246:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_84]], align 4
-// CHECK9-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_84]], align 4
-// CHECK9-NEXT:    [[SUB86:%.*]] = sub nsw i32 [[TMP247]], 0
-// CHECK9-NEXT:    [[DIV87:%.*]] = sdiv i32 [[SUB86]], 1
-// CHECK9-NEXT:    [[SUB88:%.*]] = sub nsw i32 [[DIV87]], 1
-// CHECK9-NEXT:    store i32 [[SUB88]], i32* [[DOTCAPTURE_EXPR_85]], align 4
-// CHECK9-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_85]], align 4
-// CHECK9-NEXT:    [[ADD89:%.*]] = add nsw i32 [[TMP248]], 1
-// CHECK9-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD89]] to i64
+// CHECK9-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_72]], align 4
+// CHECK9-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_72]], align 4
+// CHECK9-NEXT:    [[SUB74:%.*]] = sub nsw i32 [[TMP247]], 0
+// CHECK9-NEXT:    [[DIV75:%.*]] = sdiv i32 [[SUB74]], 1
+// CHECK9-NEXT:    [[SUB76:%.*]] = sub nsw i32 [[DIV75]], 1
+// CHECK9-NEXT:    store i32 [[SUB76]], i32* [[DOTCAPTURE_EXPR_73]], align 4
+// CHECK9-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_73]], align 4
+// CHECK9-NEXT:    [[ADD77:%.*]] = add nsw i32 [[TMP248]], 1
+// CHECK9-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD77]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP249]])
-// CHECK9-NEXT:    [[TMP250:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l562.region_id, i32 5, i8** [[TMP244]], i8** [[TMP245]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP251:%.*]] = icmp ne i32 [[TMP250]], 0
-// CHECK9-NEXT:    br i1 [[TMP251]], label [[OMP_OFFLOAD_FAILED90:%.*]], label [[OMP_OFFLOAD_CONT91:%.*]]
-// CHECK9:       omp_offload.failed90:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l562(i64 [[TMP213]], i64 [[TMP215]], double* [[TMP216]], double* [[TMP217]], double* [[TMP218]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT91]]
-// CHECK9:       omp_offload.cont91:
+// CHECK9-NEXT:    [[KERNEL_ARGS78:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP250:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP250]], align 4
+// CHECK9-NEXT:    [[TMP251:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP251]], align 4
+// CHECK9-NEXT:    [[TMP252:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP244]], i8*** [[TMP252]], align 8
+// CHECK9-NEXT:    [[TMP253:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP245]], i8*** [[TMP253]], align 8
+// CHECK9-NEXT:    [[TMP254:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP254]], align 8
+// CHECK9-NEXT:    [[TMP255:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP255]], align 8
+// CHECK9-NEXT:    [[TMP256:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP256]], align 8
+// CHECK9-NEXT:    [[TMP257:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP257]], align 8
+// CHECK9-NEXT:    [[TMP258:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l536.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]])
+// CHECK9-NEXT:    [[TMP259:%.*]] = icmp ne i32 [[TMP258]], 0
+// CHECK9-NEXT:    br i1 [[TMP259]], label [[OMP_OFFLOAD_FAILED79:%.*]], label [[OMP_OFFLOAD_CONT80:%.*]]
+// CHECK9:       omp_offload.failed79:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l536(i64 [[TMP220]], double* [[TMP221]], double* [[TMP222]], double* [[TMP223]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT80]]
+// CHECK9:       omp_offload.cont80:
+// CHECK9-NEXT:    [[TMP260:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK9-NEXT:    [[CONV82:%.*]] = bitcast i64* [[CH_CASTED81]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP260]], i32* [[CONV82]], align 4
+// CHECK9-NEXT:    [[TMP261:%.*]] = load i64, i64* [[CH_CASTED81]], align 8
+// CHECK9-NEXT:    [[TMP262:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV84:%.*]] = bitcast i64* [[N_CASTED83]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP262]], i32* [[CONV84]], align 4
+// CHECK9-NEXT:    [[TMP263:%.*]] = load i64, i64* [[N_CASTED83]], align 8
+// CHECK9-NEXT:    [[TMP264:%.*]] = load double*, double** [[A]], align 8
+// CHECK9-NEXT:    [[TMP265:%.*]] = load double*, double** [[B]], align 8
+// CHECK9-NEXT:    [[TMP266:%.*]] = load double*, double** [[C]], align 8
+// CHECK9-NEXT:    [[TMP267:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP268:%.*]] = bitcast i8** [[TMP267]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP261]], i64* [[TMP268]], align 8
+// CHECK9-NEXT:    [[TMP269:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP270:%.*]] = bitcast i8** [[TMP269]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP261]], i64* [[TMP270]], align 8
+// CHECK9-NEXT:    [[TMP271:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP271]], align 8
+// CHECK9-NEXT:    [[TMP272:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP273:%.*]] = bitcast i8** [[TMP272]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP263]], i64* [[TMP273]], align 8
+// CHECK9-NEXT:    [[TMP274:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP275:%.*]] = bitcast i8** [[TMP274]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP263]], i64* [[TMP275]], align 8
+// CHECK9-NEXT:    [[TMP276:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP276]], align 8
+// CHECK9-NEXT:    [[TMP277:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP278:%.*]] = bitcast i8** [[TMP277]] to double**
+// CHECK9-NEXT:    store double* [[TMP264]], double** [[TMP278]], align 8
+// CHECK9-NEXT:    [[TMP279:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP280:%.*]] = bitcast i8** [[TMP279]] to double**
+// CHECK9-NEXT:    store double* [[TMP264]], double** [[TMP280]], align 8
+// CHECK9-NEXT:    [[TMP281:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP281]], align 8
+// CHECK9-NEXT:    [[TMP282:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP283:%.*]] = bitcast i8** [[TMP282]] to double**
+// CHECK9-NEXT:    store double* [[TMP265]], double** [[TMP283]], align 8
+// CHECK9-NEXT:    [[TMP284:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP285:%.*]] = bitcast i8** [[TMP284]] to double**
+// CHECK9-NEXT:    store double* [[TMP265]], double** [[TMP285]], align 8
+// CHECK9-NEXT:    [[TMP286:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP286]], align 8
+// CHECK9-NEXT:    [[TMP287:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP288:%.*]] = bitcast i8** [[TMP287]] to double**
+// CHECK9-NEXT:    store double* [[TMP266]], double** [[TMP288]], align 8
+// CHECK9-NEXT:    [[TMP289:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP290:%.*]] = bitcast i8** [[TMP289]] to double**
+// CHECK9-NEXT:    store double* [[TMP266]], double** [[TMP290]], align 8
+// CHECK9-NEXT:    [[TMP291:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 4
+// CHECK9-NEXT:    store i8* null, i8** [[TMP291]], align 8
+// CHECK9-NEXT:    [[TMP292:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP293:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP294:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP294]], i32* [[DOTCAPTURE_EXPR_89]], align 4
+// CHECK9-NEXT:    [[TMP295:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_89]], align 4
+// CHECK9-NEXT:    [[SUB91:%.*]] = sub nsw i32 [[TMP295]], 0
+// CHECK9-NEXT:    [[DIV92:%.*]] = sdiv i32 [[SUB91]], 1
+// CHECK9-NEXT:    [[SUB93:%.*]] = sub nsw i32 [[DIV92]], 1
+// CHECK9-NEXT:    store i32 [[SUB93]], i32* [[DOTCAPTURE_EXPR_90]], align 4
+// CHECK9-NEXT:    [[TMP296:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_90]], align 4
+// CHECK9-NEXT:    [[ADD94:%.*]] = add nsw i32 [[TMP296]], 1
+// CHECK9-NEXT:    [[TMP297:%.*]] = zext i32 [[ADD94]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP297]])
+// CHECK9-NEXT:    [[KERNEL_ARGS95:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP298:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP298]], align 4
+// CHECK9-NEXT:    [[TMP299:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP299]], align 4
+// CHECK9-NEXT:    [[TMP300:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP292]], i8*** [[TMP300]], align 8
+// CHECK9-NEXT:    [[TMP301:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP293]], i8*** [[TMP301]], align 8
+// CHECK9-NEXT:    [[TMP302:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP302]], align 8
+// CHECK9-NEXT:    [[TMP303:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP303]], align 8
+// CHECK9-NEXT:    [[TMP304:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP304]], align 8
+// CHECK9-NEXT:    [[TMP305:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP305]], align 8
+// CHECK9-NEXT:    [[TMP306:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l562.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]])
+// CHECK9-NEXT:    [[TMP307:%.*]] = icmp ne i32 [[TMP306]], 0
+// CHECK9-NEXT:    br i1 [[TMP307]], label [[OMP_OFFLOAD_FAILED96:%.*]], label [[OMP_OFFLOAD_CONT97:%.*]]
+// CHECK9:       omp_offload.failed96:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l562(i64 [[TMP261]], i64 [[TMP263]], double* [[TMP264]], double* [[TMP265]], double* [[TMP266]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT97]]
+// CHECK9:       omp_offload.cont97:
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    ret i32 [[CALL]]
 //
@@ -6514,43 +6633,43 @@ int main() {
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[CH_CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED18:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED32:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS34:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS35:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS36:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP37:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_38:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_39:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[CH_CASTED46:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED48:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS50:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS51:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS52:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP53:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_54:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_55:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED62:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS64:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS65:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS66:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP67:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_68:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_69:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[CH_CASTED76:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED78:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS80:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS81:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS82:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP83:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_84:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_85:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED34:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS36:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS37:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS38:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP39:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_40:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_41:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[CH_CASTED49:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[N_CASTED51:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS53:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS54:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS55:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP56:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_57:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_58:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED66:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS68:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS69:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS70:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP71:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_72:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_73:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[CH_CASTED81:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[N_CASTED83:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS85:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS86:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS87:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP88:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_89:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_90:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 10000, i32* [[N]], align 4
 // CHECK9-NEXT:    store i32 100, i32* [[CH]], align 4
 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N]], align 4
@@ -6605,403 +6724,522 @@ int main() {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP29]], 1
 // CHECK9-NEXT:    [[TMP30:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP30]])
-// CHECK9-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l42.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK9-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP32]], align 4
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP25]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP26]], i8*** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l42.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK9-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l42(i64 [[TMP1]], i32* [[TMP2]], i32* [[TMP3]], i32* [[TMP4]]) #[[ATTR2]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N]], align 4
 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP33]], i32* [[CONV4]], align 4
-// CHECK9-NEXT:    [[TMP34:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK9-NEXT:    [[TMP35:%.*]] = load i32*, i32** [[A]], align 8
-// CHECK9-NEXT:    [[TMP36:%.*]] = load i32*, i32** [[B]], align 8
-// CHECK9-NEXT:    [[TMP37:%.*]] = load i32*, i32** [[C]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP41]], align 8
-// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP42]], align 8
-// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP35]], i32** [[TMP44]], align 8
-// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP35]], i32** [[TMP46]], align 8
-// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP47]], align 8
-// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP36]], i32** [[TMP49]], align 8
-// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP36]], i32** [[TMP51]], align 8
-// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP52]], align 8
-// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 3
+// CHECK9-NEXT:    store i32 [[TMP41]], i32* [[CONV4]], align 4
+// CHECK9-NEXT:    [[TMP42:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = load i32*, i32** [[A]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = load i32*, i32** [[B]], align 8
+// CHECK9-NEXT:    [[TMP45:%.*]] = load i32*, i32** [[C]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP49]], align 8
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP43]], i32** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
 // CHECK9-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP37]], i32** [[TMP54]], align 8
-// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP37]], i32** [[TMP56]], align 8
-// CHECK9-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP57]], align 8
-// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP60:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP60]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK9-NEXT:    [[TMP61:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK9-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP61]], 0
+// CHECK9-NEXT:    store i32* [[TMP43]], i32** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP44]], i32** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP44]], i32** [[TMP59]], align 8
+// CHECK9-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP60]], align 8
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP45]], i32** [[TMP62]], align 8
+// CHECK9-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP64:%.*]] = bitcast i8** [[TMP63]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP45]], i32** [[TMP64]], align 8
+// CHECK9-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP65]], align 8
+// CHECK9-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP68]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK9-NEXT:    [[TMP69:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK9-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP69]], 0
 // CHECK9-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK9-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK9-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[TMP62:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP62]], 1
-// CHECK9-NEXT:    [[TMP63:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP63]])
-// CHECK9-NEXT:    [[TMP64:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l51.region_id, i32 4, i8** [[TMP58]], i8** [[TMP59]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP65:%.*]] = icmp ne i32 [[TMP64]], 0
-// CHECK9-NEXT:    br i1 [[TMP65]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK9:       omp_offload.failed15:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l51(i64 [[TMP34]], i32* [[TMP35]], i32* [[TMP36]], i32* [[TMP37]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK9:       omp_offload.cont16:
-// CHECK9-NEXT:    [[TMP66:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK9-NEXT:    [[CONV17:%.*]] = bitcast i64* [[CH_CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP66]], i32* [[CONV17]], align 4
-// CHECK9-NEXT:    [[TMP67:%.*]] = load i64, i64* [[CH_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV19:%.*]] = bitcast i64* [[N_CASTED18]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP68]], i32* [[CONV19]], align 4
-// CHECK9-NEXT:    [[TMP69:%.*]] = load i64, i64* [[N_CASTED18]], align 8
-// CHECK9-NEXT:    [[TMP70:%.*]] = load i32*, i32** [[A]], align 8
-// CHECK9-NEXT:    [[TMP71:%.*]] = load i32*, i32** [[B]], align 8
-// CHECK9-NEXT:    [[TMP72:%.*]] = load i32*, i32** [[C]], align 8
-// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP67]], i64* [[TMP74]], align 8
-// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP67]], i64* [[TMP76]], align 8
-// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP77]], align 8
-// CHECK9-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP69]], i64* [[TMP79]], align 8
-// CHECK9-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP69]], i64* [[TMP81]], align 8
-// CHECK9-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP82]], align 8
-// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP70]], i32** [[TMP84]], align 8
-// CHECK9-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP70]], i32** [[TMP86]], align 8
-// CHECK9-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP87]], align 8
-// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP71]], i32** [[TMP89]], align 8
-// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP71]], i32** [[TMP91]], align 8
-// CHECK9-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP92]], align 8
-// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP72]], i32** [[TMP94]], align 8
-// CHECK9-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP72]], i32** [[TMP96]], align 8
-// CHECK9-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 4
-// CHECK9-NEXT:    store i8* null, i8** [[TMP97]], align 8
-// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP100]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK9-NEXT:    [[TMP101:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK9-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP101]], 0
-// CHECK9-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
-// CHECK9-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
-// CHECK9-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK9-NEXT:    [[TMP102:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK9-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP102]], 1
-// CHECK9-NEXT:    [[TMP103:%.*]] = zext i32 [[ADD29]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP103]])
-// CHECK9-NEXT:    [[TMP104:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l59.region_id, i32 5, i8** [[TMP98]], i8** [[TMP99]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP105:%.*]] = icmp ne i32 [[TMP104]], 0
-// CHECK9-NEXT:    br i1 [[TMP105]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK9:       omp_offload.failed30:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l59(i64 [[TMP67]], i64 [[TMP69]], i32* [[TMP70]], i32* [[TMP71]], i32* [[TMP72]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK9:       omp_offload.cont31:
-// CHECK9-NEXT:    [[TMP106:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV33:%.*]] = bitcast i64* [[N_CASTED32]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP106]], i32* [[CONV33]], align 4
-// CHECK9-NEXT:    [[TMP107:%.*]] = load i64, i64* [[N_CASTED32]], align 8
-// CHECK9-NEXT:    [[TMP108:%.*]] = load i32*, i32** [[A]], align 8
-// CHECK9-NEXT:    [[TMP109:%.*]] = load i32*, i32** [[B]], align 8
-// CHECK9-NEXT:    [[TMP110:%.*]] = load i32*, i32** [[C]], align 8
-// CHECK9-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP107]], i64* [[TMP112]], align 8
-// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP107]], i64* [[TMP114]], align 8
-// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP115]], align 8
-// CHECK9-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP108]], i32** [[TMP117]], align 8
-// CHECK9-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP119:%.*]] = bitcast i8** [[TMP118]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP108]], i32** [[TMP119]], align 8
-// CHECK9-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP120]], align 8
-// CHECK9-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP109]], i32** [[TMP122]], align 8
-// CHECK9-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP109]], i32** [[TMP124]], align 8
-// CHECK9-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP125]], align 8
-// CHECK9-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP110]], i32** [[TMP127]], align 8
-// CHECK9-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP110]], i32** [[TMP129]], align 8
-// CHECK9-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP130]], align 8
-// CHECK9-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP133:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP133]], i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK9-NEXT:    [[TMP134:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK9-NEXT:    [[SUB40:%.*]] = sub nsw i32 [[TMP134]], 0
-// CHECK9-NEXT:    [[DIV41:%.*]] = sdiv i32 [[SUB40]], 1
-// CHECK9-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[DIV41]], 1
-// CHECK9-NEXT:    store i32 [[SUB42]], i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK9-NEXT:    [[TMP135:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK9-NEXT:    [[ADD43:%.*]] = add nsw i32 [[TMP135]], 1
-// CHECK9-NEXT:    [[TMP136:%.*]] = zext i32 [[ADD43]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP136]])
-// CHECK9-NEXT:    [[TMP137:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l67.region_id, i32 4, i8** [[TMP131]], i8** [[TMP132]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.40, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.41, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP138:%.*]] = icmp ne i32 [[TMP137]], 0
-// CHECK9-NEXT:    br i1 [[TMP138]], label [[OMP_OFFLOAD_FAILED44:%.*]], label [[OMP_OFFLOAD_CONT45:%.*]]
-// CHECK9:       omp_offload.failed44:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l67(i64 [[TMP107]], i32* [[TMP108]], i32* [[TMP109]], i32* [[TMP110]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT45]]
-// CHECK9:       omp_offload.cont45:
-// CHECK9-NEXT:    [[TMP139:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK9-NEXT:    [[CONV47:%.*]] = bitcast i64* [[CH_CASTED46]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP139]], i32* [[CONV47]], align 4
-// CHECK9-NEXT:    [[TMP140:%.*]] = load i64, i64* [[CH_CASTED46]], align 8
-// CHECK9-NEXT:    [[TMP141:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV49:%.*]] = bitcast i64* [[N_CASTED48]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP141]], i32* [[CONV49]], align 4
-// CHECK9-NEXT:    [[TMP142:%.*]] = load i64, i64* [[N_CASTED48]], align 8
-// CHECK9-NEXT:    [[TMP143:%.*]] = load i32*, i32** [[A]], align 8
-// CHECK9-NEXT:    [[TMP144:%.*]] = load i32*, i32** [[B]], align 8
-// CHECK9-NEXT:    [[TMP145:%.*]] = load i32*, i32** [[C]], align 8
-// CHECK9-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP140]], i64* [[TMP147]], align 8
-// CHECK9-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP140]], i64* [[TMP149]], align 8
-// CHECK9-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP150]], align 8
-// CHECK9-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP142]], i64* [[TMP152]], align 8
-// CHECK9-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP142]], i64* [[TMP154]], align 8
-// CHECK9-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP155]], align 8
-// CHECK9-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP143]], i32** [[TMP157]], align 8
-// CHECK9-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP143]], i32** [[TMP159]], align 8
-// CHECK9-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP160]], align 8
-// CHECK9-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP162:%.*]] = bitcast i8** [[TMP161]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP144]], i32** [[TMP162]], align 8
-// CHECK9-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP164:%.*]] = bitcast i8** [[TMP163]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP144]], i32** [[TMP164]], align 8
-// CHECK9-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP165]], align 8
-// CHECK9-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP145]], i32** [[TMP167]], align 8
-// CHECK9-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP145]], i32** [[TMP169]], align 8
-// CHECK9-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 4
-// CHECK9-NEXT:    store i8* null, i8** [[TMP170]], align 8
-// CHECK9-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP70:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP70]], 1
+// CHECK9-NEXT:    [[TMP71:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP71]])
+// CHECK9-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP72]], align 4
+// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP73]], align 4
+// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP66]], i8*** [[TMP74]], align 8
+// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP67]], i8*** [[TMP75]], align 8
+// CHECK9-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP76]], align 8
+// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP77]], align 8
+// CHECK9-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP78]], align 8
+// CHECK9-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP79]], align 8
+// CHECK9-NEXT:    [[TMP80:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l51.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK9-NEXT:    [[TMP81:%.*]] = icmp ne i32 [[TMP80]], 0
+// CHECK9-NEXT:    br i1 [[TMP81]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK9:       omp_offload.failed16:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l51(i64 [[TMP42]], i32* [[TMP43]], i32* [[TMP44]], i32* [[TMP45]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK9:       omp_offload.cont17:
+// CHECK9-NEXT:    [[TMP82:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK9-NEXT:    [[CONV18:%.*]] = bitcast i64* [[CH_CASTED]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP82]], i32* [[CONV18]], align 4
+// CHECK9-NEXT:    [[TMP83:%.*]] = load i64, i64* [[CH_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP84:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP84]], i32* [[CONV20]], align 4
+// CHECK9-NEXT:    [[TMP85:%.*]] = load i64, i64* [[N_CASTED19]], align 8
+// CHECK9-NEXT:    [[TMP86:%.*]] = load i32*, i32** [[A]], align 8
+// CHECK9-NEXT:    [[TMP87:%.*]] = load i32*, i32** [[B]], align 8
+// CHECK9-NEXT:    [[TMP88:%.*]] = load i32*, i32** [[C]], align 8
+// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP83]], i64* [[TMP90]], align 8
+// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP83]], i64* [[TMP92]], align 8
+// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP93]], align 8
+// CHECK9-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP85]], i64* [[TMP95]], align 8
+// CHECK9-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP85]], i64* [[TMP97]], align 8
+// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP98]], align 8
+// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP100:%.*]] = bitcast i8** [[TMP99]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP86]], i32** [[TMP100]], align 8
+// CHECK9-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP86]], i32** [[TMP102]], align 8
+// CHECK9-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP103]], align 8
+// CHECK9-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP87]], i32** [[TMP105]], align 8
+// CHECK9-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP87]], i32** [[TMP107]], align 8
+// CHECK9-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP108]], align 8
+// CHECK9-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP88]], i32** [[TMP110]], align 8
+// CHECK9-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP88]], i32** [[TMP112]], align 8
+// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 4
+// CHECK9-NEXT:    store i8* null, i8** [[TMP113]], align 8
+// CHECK9-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP116:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP116]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK9-NEXT:    [[TMP117:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK9-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP117]], 0
+// CHECK9-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
+// CHECK9-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
+// CHECK9-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK9-NEXT:    [[TMP118:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK9-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP118]], 1
+// CHECK9-NEXT:    [[TMP119:%.*]] = zext i32 [[ADD30]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP119]])
+// CHECK9-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP120]], align 4
+// CHECK9-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP121]], align 4
+// CHECK9-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP114]], i8*** [[TMP122]], align 8
+// CHECK9-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP115]], i8*** [[TMP123]], align 8
+// CHECK9-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP124]], align 8
+// CHECK9-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP125]], align 8
+// CHECK9-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP126]], align 8
+// CHECK9-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP127]], align 8
+// CHECK9-NEXT:    [[TMP128:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l59.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK9-NEXT:    [[TMP129:%.*]] = icmp ne i32 [[TMP128]], 0
+// CHECK9-NEXT:    br i1 [[TMP129]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK9:       omp_offload.failed32:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l59(i64 [[TMP83]], i64 [[TMP85]], i32* [[TMP86]], i32* [[TMP87]], i32* [[TMP88]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK9:       omp_offload.cont33:
+// CHECK9-NEXT:    [[TMP130:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV35:%.*]] = bitcast i64* [[N_CASTED34]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP130]], i32* [[CONV35]], align 4
+// CHECK9-NEXT:    [[TMP131:%.*]] = load i64, i64* [[N_CASTED34]], align 8
+// CHECK9-NEXT:    [[TMP132:%.*]] = load i32*, i32** [[A]], align 8
+// CHECK9-NEXT:    [[TMP133:%.*]] = load i32*, i32** [[B]], align 8
+// CHECK9-NEXT:    [[TMP134:%.*]] = load i32*, i32** [[C]], align 8
+// CHECK9-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP131]], i64* [[TMP136]], align 8
+// CHECK9-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP131]], i64* [[TMP138]], align 8
+// CHECK9-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP139]], align 8
+// CHECK9-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP132]], i32** [[TMP141]], align 8
+// CHECK9-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP132]], i32** [[TMP143]], align 8
+// CHECK9-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP144]], align 8
+// CHECK9-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP133]], i32** [[TMP146]], align 8
+// CHECK9-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP133]], i32** [[TMP148]], align 8
+// CHECK9-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP149]], align 8
+// CHECK9-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP134]], i32** [[TMP151]], align 8
+// CHECK9-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP134]], i32** [[TMP153]], align 8
+// CHECK9-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP154]], align 8
+// CHECK9-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP157:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP157]], i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK9-NEXT:    [[TMP158:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK9-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[TMP158]], 0
+// CHECK9-NEXT:    [[DIV43:%.*]] = sdiv i32 [[SUB42]], 1
+// CHECK9-NEXT:    [[SUB44:%.*]] = sub nsw i32 [[DIV43]], 1
+// CHECK9-NEXT:    store i32 [[SUB44]], i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK9-NEXT:    [[TMP159:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK9-NEXT:    [[ADD45:%.*]] = add nsw i32 [[TMP159]], 1
+// CHECK9-NEXT:    [[TMP160:%.*]] = zext i32 [[ADD45]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP160]])
+// CHECK9-NEXT:    [[KERNEL_ARGS46:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP161]], align 4
+// CHECK9-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP162]], align 4
+// CHECK9-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP155]], i8*** [[TMP163]], align 8
+// CHECK9-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP156]], i8*** [[TMP164]], align 8
+// CHECK9-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.40, i32 0, i32 0), i64** [[TMP165]], align 8
+// CHECK9-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.41, i32 0, i32 0), i64** [[TMP166]], align 8
+// CHECK9-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP167]], align 8
+// CHECK9-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP168]], align 8
+// CHECK9-NEXT:    [[TMP169:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l67.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]])
+// CHECK9-NEXT:    [[TMP170:%.*]] = icmp ne i32 [[TMP169]], 0
+// CHECK9-NEXT:    br i1 [[TMP170]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
+// CHECK9:       omp_offload.failed47:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l67(i64 [[TMP131]], i32* [[TMP132]], i32* [[TMP133]], i32* [[TMP134]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
+// CHECK9:       omp_offload.cont48:
+// CHECK9-NEXT:    [[TMP171:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK9-NEXT:    [[CONV50:%.*]] = bitcast i64* [[CH_CASTED49]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP171]], i32* [[CONV50]], align 4
+// CHECK9-NEXT:    [[TMP172:%.*]] = load i64, i64* [[CH_CASTED49]], align 8
 // CHECK9-NEXT:    [[TMP173:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP173]], i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK9-NEXT:    [[TMP174:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK9-NEXT:    [[SUB56:%.*]] = sub nsw i32 [[TMP174]], 0
-// CHECK9-NEXT:    [[DIV57:%.*]] = sdiv i32 [[SUB56]], 1
-// CHECK9-NEXT:    [[SUB58:%.*]] = sub nsw i32 [[DIV57]], 1
-// CHECK9-NEXT:    store i32 [[SUB58]], i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK9-NEXT:    [[TMP175:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK9-NEXT:    [[ADD59:%.*]] = add nsw i32 [[TMP175]], 1
-// CHECK9-NEXT:    [[TMP176:%.*]] = zext i32 [[ADD59]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP176]])
-// CHECK9-NEXT:    [[TMP177:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75.region_id, i32 5, i8** [[TMP171]], i8** [[TMP172]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.44, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.45, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP178:%.*]] = icmp ne i32 [[TMP177]], 0
-// CHECK9-NEXT:    br i1 [[TMP178]], label [[OMP_OFFLOAD_FAILED60:%.*]], label [[OMP_OFFLOAD_CONT61:%.*]]
-// CHECK9:       omp_offload.failed60:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75(i64 [[TMP140]], i64 [[TMP142]], i32* [[TMP143]], i32* [[TMP144]], i32* [[TMP145]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT61]]
-// CHECK9:       omp_offload.cont61:
-// CHECK9-NEXT:    [[TMP179:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV63:%.*]] = bitcast i64* [[N_CASTED62]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP179]], i32* [[CONV63]], align 4
-// CHECK9-NEXT:    [[TMP180:%.*]] = load i64, i64* [[N_CASTED62]], align 8
-// CHECK9-NEXT:    [[TMP181:%.*]] = load i32*, i32** [[A]], align 8
-// CHECK9-NEXT:    [[TMP182:%.*]] = load i32*, i32** [[B]], align 8
-// CHECK9-NEXT:    [[TMP183:%.*]] = load i32*, i32** [[C]], align 8
-// CHECK9-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP185:%.*]] = bitcast i8** [[TMP184]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP180]], i64* [[TMP185]], align 8
-// CHECK9-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP187:%.*]] = bitcast i8** [[TMP186]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP180]], i64* [[TMP187]], align 8
-// CHECK9-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP188]], align 8
-// CHECK9-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP190:%.*]] = bitcast i8** [[TMP189]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP181]], i32** [[TMP190]], align 8
-// CHECK9-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP192:%.*]] = bitcast i8** [[TMP191]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP181]], i32** [[TMP192]], align 8
-// CHECK9-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP193]], align 8
-// CHECK9-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP195:%.*]] = bitcast i8** [[TMP194]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP182]], i32** [[TMP195]], align 8
-// CHECK9-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP197:%.*]] = bitcast i8** [[TMP196]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP182]], i32** [[TMP197]], align 8
-// CHECK9-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP198]], align 8
-// CHECK9-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP200:%.*]] = bitcast i8** [[TMP199]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP183]], i32** [[TMP200]], align 8
-// CHECK9-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP202:%.*]] = bitcast i8** [[TMP201]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP183]], i32** [[TMP202]], align 8
-// CHECK9-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP203]], align 8
-// CHECK9-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP206:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP206]], i32* [[DOTCAPTURE_EXPR_68]], align 4
-// CHECK9-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_68]], align 4
-// CHECK9-NEXT:    [[SUB70:%.*]] = sub nsw i32 [[TMP207]], 0
-// CHECK9-NEXT:    [[DIV71:%.*]] = sdiv i32 [[SUB70]], 1
-// CHECK9-NEXT:    [[SUB72:%.*]] = sub nsw i32 [[DIV71]], 1
-// CHECK9-NEXT:    store i32 [[SUB72]], i32* [[DOTCAPTURE_EXPR_69]], align 4
-// CHECK9-NEXT:    [[TMP208:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_69]], align 4
-// CHECK9-NEXT:    [[ADD73:%.*]] = add nsw i32 [[TMP208]], 1
-// CHECK9-NEXT:    [[TMP209:%.*]] = zext i32 [[ADD73]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP209]])
-// CHECK9-NEXT:    [[TMP210:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l83.region_id, i32 4, i8** [[TMP204]], i8** [[TMP205]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.48, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.49, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP211:%.*]] = icmp ne i32 [[TMP210]], 0
-// CHECK9-NEXT:    br i1 [[TMP211]], label [[OMP_OFFLOAD_FAILED74:%.*]], label [[OMP_OFFLOAD_CONT75:%.*]]
-// CHECK9:       omp_offload.failed74:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l83(i64 [[TMP180]], i32* [[TMP181]], i32* [[TMP182]], i32* [[TMP183]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT75]]
-// CHECK9:       omp_offload.cont75:
-// CHECK9-NEXT:    [[TMP212:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK9-NEXT:    [[CONV77:%.*]] = bitcast i64* [[CH_CASTED76]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP212]], i32* [[CONV77]], align 4
-// CHECK9-NEXT:    [[TMP213:%.*]] = load i64, i64* [[CH_CASTED76]], align 8
-// CHECK9-NEXT:    [[TMP214:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV79:%.*]] = bitcast i64* [[N_CASTED78]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP214]], i32* [[CONV79]], align 4
-// CHECK9-NEXT:    [[TMP215:%.*]] = load i64, i64* [[N_CASTED78]], align 8
-// CHECK9-NEXT:    [[TMP216:%.*]] = load i32*, i32** [[A]], align 8
-// CHECK9-NEXT:    [[TMP217:%.*]] = load i32*, i32** [[B]], align 8
-// CHECK9-NEXT:    [[TMP218:%.*]] = load i32*, i32** [[C]], align 8
-// CHECK9-NEXT:    [[TMP219:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP220:%.*]] = bitcast i8** [[TMP219]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP213]], i64* [[TMP220]], align 8
-// CHECK9-NEXT:    [[TMP221:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP222:%.*]] = bitcast i8** [[TMP221]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP213]], i64* [[TMP222]], align 8
-// CHECK9-NEXT:    [[TMP223:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP223]], align 8
-// CHECK9-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 1
+// CHECK9-NEXT:    [[CONV52:%.*]] = bitcast i64* [[N_CASTED51]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP173]], i32* [[CONV52]], align 4
+// CHECK9-NEXT:    [[TMP174:%.*]] = load i64, i64* [[N_CASTED51]], align 8
+// CHECK9-NEXT:    [[TMP175:%.*]] = load i32*, i32** [[A]], align 8
+// CHECK9-NEXT:    [[TMP176:%.*]] = load i32*, i32** [[B]], align 8
+// CHECK9-NEXT:    [[TMP177:%.*]] = load i32*, i32** [[C]], align 8
+// CHECK9-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP172]], i64* [[TMP179]], align 8
+// CHECK9-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP172]], i64* [[TMP181]], align 8
+// CHECK9-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP182]], align 8
+// CHECK9-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP184:%.*]] = bitcast i8** [[TMP183]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP174]], i64* [[TMP184]], align 8
+// CHECK9-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP186:%.*]] = bitcast i8** [[TMP185]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP174]], i64* [[TMP186]], align 8
+// CHECK9-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP187]], align 8
+// CHECK9-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP189:%.*]] = bitcast i8** [[TMP188]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP175]], i32** [[TMP189]], align 8
+// CHECK9-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP191:%.*]] = bitcast i8** [[TMP190]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP175]], i32** [[TMP191]], align 8
+// CHECK9-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP192]], align 8
+// CHECK9-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP194:%.*]] = bitcast i8** [[TMP193]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP176]], i32** [[TMP194]], align 8
+// CHECK9-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP196:%.*]] = bitcast i8** [[TMP195]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP176]], i32** [[TMP196]], align 8
+// CHECK9-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP197]], align 8
+// CHECK9-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP199:%.*]] = bitcast i8** [[TMP198]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP177]], i32** [[TMP199]], align 8
+// CHECK9-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP201:%.*]] = bitcast i8** [[TMP200]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP177]], i32** [[TMP201]], align 8
+// CHECK9-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 4
+// CHECK9-NEXT:    store i8* null, i8** [[TMP202]], align 8
+// CHECK9-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP205:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP205]], i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK9-NEXT:    [[TMP206:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK9-NEXT:    [[SUB59:%.*]] = sub nsw i32 [[TMP206]], 0
+// CHECK9-NEXT:    [[DIV60:%.*]] = sdiv i32 [[SUB59]], 1
+// CHECK9-NEXT:    [[SUB61:%.*]] = sub nsw i32 [[DIV60]], 1
+// CHECK9-NEXT:    store i32 [[SUB61]], i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK9-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK9-NEXT:    [[ADD62:%.*]] = add nsw i32 [[TMP207]], 1
+// CHECK9-NEXT:    [[TMP208:%.*]] = zext i32 [[ADD62]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP208]])
+// CHECK9-NEXT:    [[KERNEL_ARGS63:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP209:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP209]], align 4
+// CHECK9-NEXT:    [[TMP210:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP210]], align 4
+// CHECK9-NEXT:    [[TMP211:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP203]], i8*** [[TMP211]], align 8
+// CHECK9-NEXT:    [[TMP212:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP204]], i8*** [[TMP212]], align 8
+// CHECK9-NEXT:    [[TMP213:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.44, i32 0, i32 0), i64** [[TMP213]], align 8
+// CHECK9-NEXT:    [[TMP214:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.45, i32 0, i32 0), i64** [[TMP214]], align 8
+// CHECK9-NEXT:    [[TMP215:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP215]], align 8
+// CHECK9-NEXT:    [[TMP216:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP216]], align 8
+// CHECK9-NEXT:    [[TMP217:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]])
+// CHECK9-NEXT:    [[TMP218:%.*]] = icmp ne i32 [[TMP217]], 0
+// CHECK9-NEXT:    br i1 [[TMP218]], label [[OMP_OFFLOAD_FAILED64:%.*]], label [[OMP_OFFLOAD_CONT65:%.*]]
+// CHECK9:       omp_offload.failed64:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75(i64 [[TMP172]], i64 [[TMP174]], i32* [[TMP175]], i32* [[TMP176]], i32* [[TMP177]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT65]]
+// CHECK9:       omp_offload.cont65:
+// CHECK9-NEXT:    [[TMP219:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV67:%.*]] = bitcast i64* [[N_CASTED66]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP219]], i32* [[CONV67]], align 4
+// CHECK9-NEXT:    [[TMP220:%.*]] = load i64, i64* [[N_CASTED66]], align 8
+// CHECK9-NEXT:    [[TMP221:%.*]] = load i32*, i32** [[A]], align 8
+// CHECK9-NEXT:    [[TMP222:%.*]] = load i32*, i32** [[B]], align 8
+// CHECK9-NEXT:    [[TMP223:%.*]] = load i32*, i32** [[C]], align 8
+// CHECK9-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP225:%.*]] = bitcast i8** [[TMP224]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP215]], i64* [[TMP225]], align 8
-// CHECK9-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 1
+// CHECK9-NEXT:    store i64 [[TMP220]], i64* [[TMP225]], align 8
+// CHECK9-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP227:%.*]] = bitcast i8** [[TMP226]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP215]], i64* [[TMP227]], align 8
-// CHECK9-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 1
+// CHECK9-NEXT:    store i64 [[TMP220]], i64* [[TMP227]], align 8
+// CHECK9-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 0
 // CHECK9-NEXT:    store i8* null, i8** [[TMP228]], align 8
-// CHECK9-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 1
 // CHECK9-NEXT:    [[TMP230:%.*]] = bitcast i8** [[TMP229]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP216]], i32** [[TMP230]], align 8
-// CHECK9-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 2
+// CHECK9-NEXT:    store i32* [[TMP221]], i32** [[TMP230]], align 8
+// CHECK9-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 1
 // CHECK9-NEXT:    [[TMP232:%.*]] = bitcast i8** [[TMP231]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP216]], i32** [[TMP232]], align 8
-// CHECK9-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 2
+// CHECK9-NEXT:    store i32* [[TMP221]], i32** [[TMP232]], align 8
+// CHECK9-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 1
 // CHECK9-NEXT:    store i8* null, i8** [[TMP233]], align 8
-// CHECK9-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 2
 // CHECK9-NEXT:    [[TMP235:%.*]] = bitcast i8** [[TMP234]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP217]], i32** [[TMP235]], align 8
-// CHECK9-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 3
+// CHECK9-NEXT:    store i32* [[TMP222]], i32** [[TMP235]], align 8
+// CHECK9-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 2
 // CHECK9-NEXT:    [[TMP237:%.*]] = bitcast i8** [[TMP236]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP217]], i32** [[TMP237]], align 8
-// CHECK9-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 3
+// CHECK9-NEXT:    store i32* [[TMP222]], i32** [[TMP237]], align 8
+// CHECK9-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 2
 // CHECK9-NEXT:    store i8* null, i8** [[TMP238]], align 8
-// CHECK9-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 3
 // CHECK9-NEXT:    [[TMP240:%.*]] = bitcast i8** [[TMP239]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP218]], i32** [[TMP240]], align 8
-// CHECK9-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 4
+// CHECK9-NEXT:    store i32* [[TMP223]], i32** [[TMP240]], align 8
+// CHECK9-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 3
 // CHECK9-NEXT:    [[TMP242:%.*]] = bitcast i8** [[TMP241]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP218]], i32** [[TMP242]], align 8
-// CHECK9-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 4
+// CHECK9-NEXT:    store i32* [[TMP223]], i32** [[TMP242]], align 8
+// CHECK9-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 3
 // CHECK9-NEXT:    store i8* null, i8** [[TMP243]], align 8
-// CHECK9-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP246:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_84]], align 4
-// CHECK9-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_84]], align 4
-// CHECK9-NEXT:    [[SUB86:%.*]] = sub nsw i32 [[TMP247]], 0
-// CHECK9-NEXT:    [[DIV87:%.*]] = sdiv i32 [[SUB86]], 1
-// CHECK9-NEXT:    [[SUB88:%.*]] = sub nsw i32 [[DIV87]], 1
-// CHECK9-NEXT:    store i32 [[SUB88]], i32* [[DOTCAPTURE_EXPR_85]], align 4
-// CHECK9-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_85]], align 4
-// CHECK9-NEXT:    [[ADD89:%.*]] = add nsw i32 [[TMP248]], 1
-// CHECK9-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD89]] to i64
+// CHECK9-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_72]], align 4
+// CHECK9-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_72]], align 4
+// CHECK9-NEXT:    [[SUB74:%.*]] = sub nsw i32 [[TMP247]], 0
+// CHECK9-NEXT:    [[DIV75:%.*]] = sdiv i32 [[SUB74]], 1
+// CHECK9-NEXT:    [[SUB76:%.*]] = sub nsw i32 [[DIV75]], 1
+// CHECK9-NEXT:    store i32 [[SUB76]], i32* [[DOTCAPTURE_EXPR_73]], align 4
+// CHECK9-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_73]], align 4
+// CHECK9-NEXT:    [[ADD77:%.*]] = add nsw i32 [[TMP248]], 1
+// CHECK9-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD77]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP249]])
-// CHECK9-NEXT:    [[TMP250:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l91.region_id, i32 5, i8** [[TMP244]], i8** [[TMP245]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.52, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.53, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP251:%.*]] = icmp ne i32 [[TMP250]], 0
-// CHECK9-NEXT:    br i1 [[TMP251]], label [[OMP_OFFLOAD_FAILED90:%.*]], label [[OMP_OFFLOAD_CONT91:%.*]]
-// CHECK9:       omp_offload.failed90:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l91(i64 [[TMP213]], i64 [[TMP215]], i32* [[TMP216]], i32* [[TMP217]], i32* [[TMP218]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT91]]
-// CHECK9:       omp_offload.cont91:
+// CHECK9-NEXT:    [[KERNEL_ARGS78:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP250:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP250]], align 4
+// CHECK9-NEXT:    [[TMP251:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP251]], align 4
+// CHECK9-NEXT:    [[TMP252:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP244]], i8*** [[TMP252]], align 8
+// CHECK9-NEXT:    [[TMP253:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP245]], i8*** [[TMP253]], align 8
+// CHECK9-NEXT:    [[TMP254:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.48, i32 0, i32 0), i64** [[TMP254]], align 8
+// CHECK9-NEXT:    [[TMP255:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.49, i32 0, i32 0), i64** [[TMP255]], align 8
+// CHECK9-NEXT:    [[TMP256:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP256]], align 8
+// CHECK9-NEXT:    [[TMP257:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP257]], align 8
+// CHECK9-NEXT:    [[TMP258:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l83.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]])
+// CHECK9-NEXT:    [[TMP259:%.*]] = icmp ne i32 [[TMP258]], 0
+// CHECK9-NEXT:    br i1 [[TMP259]], label [[OMP_OFFLOAD_FAILED79:%.*]], label [[OMP_OFFLOAD_CONT80:%.*]]
+// CHECK9:       omp_offload.failed79:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l83(i64 [[TMP220]], i32* [[TMP221]], i32* [[TMP222]], i32* [[TMP223]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT80]]
+// CHECK9:       omp_offload.cont80:
+// CHECK9-NEXT:    [[TMP260:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK9-NEXT:    [[CONV82:%.*]] = bitcast i64* [[CH_CASTED81]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP260]], i32* [[CONV82]], align 4
+// CHECK9-NEXT:    [[TMP261:%.*]] = load i64, i64* [[CH_CASTED81]], align 8
+// CHECK9-NEXT:    [[TMP262:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV84:%.*]] = bitcast i64* [[N_CASTED83]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP262]], i32* [[CONV84]], align 4
+// CHECK9-NEXT:    [[TMP263:%.*]] = load i64, i64* [[N_CASTED83]], align 8
+// CHECK9-NEXT:    [[TMP264:%.*]] = load i32*, i32** [[A]], align 8
+// CHECK9-NEXT:    [[TMP265:%.*]] = load i32*, i32** [[B]], align 8
+// CHECK9-NEXT:    [[TMP266:%.*]] = load i32*, i32** [[C]], align 8
+// CHECK9-NEXT:    [[TMP267:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP268:%.*]] = bitcast i8** [[TMP267]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP261]], i64* [[TMP268]], align 8
+// CHECK9-NEXT:    [[TMP269:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP270:%.*]] = bitcast i8** [[TMP269]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP261]], i64* [[TMP270]], align 8
+// CHECK9-NEXT:    [[TMP271:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP271]], align 8
+// CHECK9-NEXT:    [[TMP272:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP273:%.*]] = bitcast i8** [[TMP272]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP263]], i64* [[TMP273]], align 8
+// CHECK9-NEXT:    [[TMP274:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP275:%.*]] = bitcast i8** [[TMP274]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP263]], i64* [[TMP275]], align 8
+// CHECK9-NEXT:    [[TMP276:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP276]], align 8
+// CHECK9-NEXT:    [[TMP277:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP278:%.*]] = bitcast i8** [[TMP277]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP264]], i32** [[TMP278]], align 8
+// CHECK9-NEXT:    [[TMP279:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP280:%.*]] = bitcast i8** [[TMP279]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP264]], i32** [[TMP280]], align 8
+// CHECK9-NEXT:    [[TMP281:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP281]], align 8
+// CHECK9-NEXT:    [[TMP282:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP283:%.*]] = bitcast i8** [[TMP282]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP265]], i32** [[TMP283]], align 8
+// CHECK9-NEXT:    [[TMP284:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP285:%.*]] = bitcast i8** [[TMP284]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP265]], i32** [[TMP285]], align 8
+// CHECK9-NEXT:    [[TMP286:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP286]], align 8
+// CHECK9-NEXT:    [[TMP287:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP288:%.*]] = bitcast i8** [[TMP287]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP266]], i32** [[TMP288]], align 8
+// CHECK9-NEXT:    [[TMP289:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP290:%.*]] = bitcast i8** [[TMP289]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP266]], i32** [[TMP290]], align 8
+// CHECK9-NEXT:    [[TMP291:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 4
+// CHECK9-NEXT:    store i8* null, i8** [[TMP291]], align 8
+// CHECK9-NEXT:    [[TMP292:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP293:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP294:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP294]], i32* [[DOTCAPTURE_EXPR_89]], align 4
+// CHECK9-NEXT:    [[TMP295:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_89]], align 4
+// CHECK9-NEXT:    [[SUB91:%.*]] = sub nsw i32 [[TMP295]], 0
+// CHECK9-NEXT:    [[DIV92:%.*]] = sdiv i32 [[SUB91]], 1
+// CHECK9-NEXT:    [[SUB93:%.*]] = sub nsw i32 [[DIV92]], 1
+// CHECK9-NEXT:    store i32 [[SUB93]], i32* [[DOTCAPTURE_EXPR_90]], align 4
+// CHECK9-NEXT:    [[TMP296:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_90]], align 4
+// CHECK9-NEXT:    [[ADD94:%.*]] = add nsw i32 [[TMP296]], 1
+// CHECK9-NEXT:    [[TMP297:%.*]] = zext i32 [[ADD94]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP297]])
+// CHECK9-NEXT:    [[KERNEL_ARGS95:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP298:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP298]], align 4
+// CHECK9-NEXT:    [[TMP299:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP299]], align 4
+// CHECK9-NEXT:    [[TMP300:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP292]], i8*** [[TMP300]], align 8
+// CHECK9-NEXT:    [[TMP301:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP293]], i8*** [[TMP301]], align 8
+// CHECK9-NEXT:    [[TMP302:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.52, i32 0, i32 0), i64** [[TMP302]], align 8
+// CHECK9-NEXT:    [[TMP303:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.53, i32 0, i32 0), i64** [[TMP303]], align 8
+// CHECK9-NEXT:    [[TMP304:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP304]], align 8
+// CHECK9-NEXT:    [[TMP305:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP305]], align 8
+// CHECK9-NEXT:    [[TMP306:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l91.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]])
+// CHECK9-NEXT:    [[TMP307:%.*]] = icmp ne i32 [[TMP306]], 0
+// CHECK9-NEXT:    br i1 [[TMP307]], label [[OMP_OFFLOAD_FAILED96:%.*]], label [[OMP_OFFLOAD_CONT97:%.*]]
+// CHECK9:       omp_offload.failed96:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l91(i64 [[TMP261]], i64 [[TMP263]], i32* [[TMP264]], i32* [[TMP265]], i32* [[TMP266]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT97]]
+// CHECK9:       omp_offload.cont97:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -8739,43 +8977,43 @@ int main() {
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_8:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[CH_CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED16:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_21:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED17:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS18:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS19:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS20:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP21:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED29:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS30:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS31:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS32:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP33:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_34:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_35:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[CH_CASTED42:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED43:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS44:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS45:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS46:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP47:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_48:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_49:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED56:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS57:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS58:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS59:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP60:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_61:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_62:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[CH_CASTED69:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED70:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS71:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS72:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS73:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP74:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_75:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_76:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_23:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED31:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS32:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS33:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS34:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP35:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_36:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_37:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[CH_CASTED45:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED46:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS47:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS48:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS49:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP50:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_51:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_52:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED60:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS61:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS62:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS63:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP64:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_65:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_66:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[CH_CASTED74:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED75:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS76:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS77:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS78:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP79:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_80:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_81:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    store i32 10000, i32* [[N]], align 4
 // CHECK11-NEXT:    store i32 100, i32* [[CH]], align 4
@@ -8830,394 +9068,513 @@ int main() {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP29]], 1
 // CHECK11-NEXT:    [[TMP30:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP30]])
-// CHECK11-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l369.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK11-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP25]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP26]], i8*** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l369.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK11-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l369(i32 [[TMP1]], double* [[TMP2]], double* [[TMP3]], double* [[TMP4]]) #[[ATTR2:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP34:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = load double*, double** [[A]], align 4
-// CHECK11-NEXT:    [[TMP36:%.*]] = load double*, double** [[B]], align 4
-// CHECK11-NEXT:    [[TMP37:%.*]] = load double*, double** [[C]], align 4
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP34]], i32* [[TMP39]], align 4
-// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP34]], i32* [[TMP41]], align 4
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP42]], align 4
-// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to double**
-// CHECK11-NEXT:    store double* [[TMP35]], double** [[TMP44]], align 4
-// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to double**
-// CHECK11-NEXT:    store double* [[TMP35]], double** [[TMP46]], align 4
-// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP47]], align 4
-// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to double**
-// CHECK11-NEXT:    store double* [[TMP36]], double** [[TMP49]], align 4
-// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to double**
-// CHECK11-NEXT:    store double* [[TMP36]], double** [[TMP51]], align 4
-// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = load double*, double** [[A]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = load double*, double** [[B]], align 4
+// CHECK11-NEXT:    [[TMP45:%.*]] = load double*, double** [[C]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP42]], i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP42]], i32* [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to double**
+// CHECK11-NEXT:    store double* [[TMP43]], double** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
 // CHECK11-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to double**
-// CHECK11-NEXT:    store double* [[TMP37]], double** [[TMP54]], align 4
-// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to double**
-// CHECK11-NEXT:    store double* [[TMP37]], double** [[TMP56]], align 4
-// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP57]], align 4
-// CHECK11-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP60:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP60]], i32* [[DOTCAPTURE_EXPR_8]], align 4
-// CHECK11-NEXT:    [[TMP61:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_8]], align 4
-// CHECK11-NEXT:    [[SUB10:%.*]] = sub nsw i32 [[TMP61]], 0
+// CHECK11-NEXT:    store double* [[TMP43]], double** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP55]], align 4
+// CHECK11-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to double**
+// CHECK11-NEXT:    store double* [[TMP44]], double** [[TMP57]], align 4
+// CHECK11-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to double**
+// CHECK11-NEXT:    store double* [[TMP44]], double** [[TMP59]], align 4
+// CHECK11-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to double**
+// CHECK11-NEXT:    store double* [[TMP45]], double** [[TMP62]], align 4
+// CHECK11-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP64:%.*]] = bitcast i8** [[TMP63]] to double**
+// CHECK11-NEXT:    store double* [[TMP45]], double** [[TMP64]], align 4
+// CHECK11-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP65]], align 4
+// CHECK11-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP68]], i32* [[DOTCAPTURE_EXPR_8]], align 4
+// CHECK11-NEXT:    [[TMP69:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_8]], align 4
+// CHECK11-NEXT:    [[SUB10:%.*]] = sub nsw i32 [[TMP69]], 0
 // CHECK11-NEXT:    [[DIV11:%.*]] = sdiv i32 [[SUB10]], 1
 // CHECK11-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[DIV11]], 1
 // CHECK11-NEXT:    store i32 [[SUB12]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[TMP62:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[ADD13:%.*]] = add nsw i32 [[TMP62]], 1
-// CHECK11-NEXT:    [[TMP63:%.*]] = zext i32 [[ADD13]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP63]])
-// CHECK11-NEXT:    [[TMP64:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l408.region_id, i32 4, i8** [[TMP58]], i8** [[TMP59]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP65:%.*]] = icmp ne i32 [[TMP64]], 0
-// CHECK11-NEXT:    br i1 [[TMP65]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK11:       omp_offload.failed14:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l408(i32 [[TMP34]], double* [[TMP35]], double* [[TMP36]], double* [[TMP37]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK11:       omp_offload.cont15:
-// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK11-NEXT:    store i32 [[TMP66]], i32* [[CH_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP67:%.*]] = load i32, i32* [[CH_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP68]], i32* [[N_CASTED16]], align 4
-// CHECK11-NEXT:    [[TMP69:%.*]] = load i32, i32* [[N_CASTED16]], align 4
-// CHECK11-NEXT:    [[TMP70:%.*]] = load double*, double** [[A]], align 4
-// CHECK11-NEXT:    [[TMP71:%.*]] = load double*, double** [[B]], align 4
-// CHECK11-NEXT:    [[TMP72:%.*]] = load double*, double** [[C]], align 4
-// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP67]], i32* [[TMP74]], align 4
-// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP67]], i32* [[TMP76]], align 4
-// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP77]], align 4
-// CHECK11-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP69]], i32* [[TMP79]], align 4
-// CHECK11-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP69]], i32* [[TMP81]], align 4
-// CHECK11-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP82]], align 4
-// CHECK11-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to double**
-// CHECK11-NEXT:    store double* [[TMP70]], double** [[TMP84]], align 4
-// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to double**
-// CHECK11-NEXT:    store double* [[TMP70]], double** [[TMP86]], align 4
-// CHECK11-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP87]], align 4
-// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to double**
-// CHECK11-NEXT:    store double* [[TMP71]], double** [[TMP89]], align 4
-// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to double**
-// CHECK11-NEXT:    store double* [[TMP71]], double** [[TMP91]], align 4
-// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP92]], align 4
-// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to double**
-// CHECK11-NEXT:    store double* [[TMP72]], double** [[TMP94]], align 4
-// CHECK11-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to double**
-// CHECK11-NEXT:    store double* [[TMP72]], double** [[TMP96]], align 4
-// CHECK11-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 4
-// CHECK11-NEXT:    store i8* null, i8** [[TMP97]], align 4
-// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP100]], i32* [[DOTCAPTURE_EXPR_21]], align 4
-// CHECK11-NEXT:    [[TMP101:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_21]], align 4
-// CHECK11-NEXT:    [[SUB23:%.*]] = sub nsw i32 [[TMP101]], 0
-// CHECK11-NEXT:    [[DIV24:%.*]] = sdiv i32 [[SUB23]], 1
-// CHECK11-NEXT:    [[SUB25:%.*]] = sub nsw i32 [[DIV24]], 1
-// CHECK11-NEXT:    store i32 [[SUB25]], i32* [[DOTCAPTURE_EXPR_22]], align 4
-// CHECK11-NEXT:    [[TMP102:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
-// CHECK11-NEXT:    [[ADD26:%.*]] = add nsw i32 [[TMP102]], 1
-// CHECK11-NEXT:    [[TMP103:%.*]] = zext i32 [[ADD26]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP103]])
-// CHECK11-NEXT:    [[TMP104:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l447.region_id, i32 5, i8** [[TMP98]], i8** [[TMP99]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP105:%.*]] = icmp ne i32 [[TMP104]], 0
-// CHECK11-NEXT:    br i1 [[TMP105]], label [[OMP_OFFLOAD_FAILED27:%.*]], label [[OMP_OFFLOAD_CONT28:%.*]]
-// CHECK11:       omp_offload.failed27:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l447(i32 [[TMP67]], i32 [[TMP69]], double* [[TMP70]], double* [[TMP71]], double* [[TMP72]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT28]]
-// CHECK11:       omp_offload.cont28:
-// CHECK11-NEXT:    [[TMP106:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP106]], i32* [[N_CASTED29]], align 4
-// CHECK11-NEXT:    [[TMP107:%.*]] = load i32, i32* [[N_CASTED29]], align 4
-// CHECK11-NEXT:    [[TMP108:%.*]] = load double*, double** [[A]], align 4
-// CHECK11-NEXT:    [[TMP109:%.*]] = load double*, double** [[B]], align 4
-// CHECK11-NEXT:    [[TMP110:%.*]] = load double*, double** [[C]], align 4
-// CHECK11-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP107]], i32* [[TMP112]], align 4
-// CHECK11-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP107]], i32* [[TMP114]], align 4
-// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP115]], align 4
-// CHECK11-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to double**
-// CHECK11-NEXT:    store double* [[TMP108]], double** [[TMP117]], align 4
-// CHECK11-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP119:%.*]] = bitcast i8** [[TMP118]] to double**
-// CHECK11-NEXT:    store double* [[TMP108]], double** [[TMP119]], align 4
-// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP120]], align 4
-// CHECK11-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to double**
-// CHECK11-NEXT:    store double* [[TMP109]], double** [[TMP122]], align 4
-// CHECK11-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to double**
-// CHECK11-NEXT:    store double* [[TMP109]], double** [[TMP124]], align 4
-// CHECK11-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP125]], align 4
-// CHECK11-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to double**
-// CHECK11-NEXT:    store double* [[TMP110]], double** [[TMP127]], align 4
-// CHECK11-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to double**
-// CHECK11-NEXT:    store double* [[TMP110]], double** [[TMP129]], align 4
-// CHECK11-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP130]], align 4
-// CHECK11-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP133:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP133]], i32* [[DOTCAPTURE_EXPR_34]], align 4
-// CHECK11-NEXT:    [[TMP134:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_34]], align 4
-// CHECK11-NEXT:    [[SUB36:%.*]] = sub nsw i32 [[TMP134]], 0
-// CHECK11-NEXT:    [[DIV37:%.*]] = sdiv i32 [[SUB36]], 1
-// CHECK11-NEXT:    [[SUB38:%.*]] = sub nsw i32 [[DIV37]], 1
-// CHECK11-NEXT:    store i32 [[SUB38]], i32* [[DOTCAPTURE_EXPR_35]], align 4
-// CHECK11-NEXT:    [[TMP135:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_35]], align 4
-// CHECK11-NEXT:    [[ADD39:%.*]] = add nsw i32 [[TMP135]], 1
-// CHECK11-NEXT:    [[TMP136:%.*]] = zext i32 [[ADD39]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP136]])
-// CHECK11-NEXT:    [[TMP137:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l478.region_id, i32 4, i8** [[TMP131]], i8** [[TMP132]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP138:%.*]] = icmp ne i32 [[TMP137]], 0
-// CHECK11-NEXT:    br i1 [[TMP138]], label [[OMP_OFFLOAD_FAILED40:%.*]], label [[OMP_OFFLOAD_CONT41:%.*]]
-// CHECK11:       omp_offload.failed40:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l478(i32 [[TMP107]], double* [[TMP108]], double* [[TMP109]], double* [[TMP110]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT41]]
-// CHECK11:       omp_offload.cont41:
-// CHECK11-NEXT:    [[TMP139:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK11-NEXT:    store i32 [[TMP139]], i32* [[CH_CASTED42]], align 4
-// CHECK11-NEXT:    [[TMP140:%.*]] = load i32, i32* [[CH_CASTED42]], align 4
-// CHECK11-NEXT:    [[TMP141:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP141]], i32* [[N_CASTED43]], align 4
-// CHECK11-NEXT:    [[TMP142:%.*]] = load i32, i32* [[N_CASTED43]], align 4
-// CHECK11-NEXT:    [[TMP143:%.*]] = load double*, double** [[A]], align 4
-// CHECK11-NEXT:    [[TMP144:%.*]] = load double*, double** [[B]], align 4
-// CHECK11-NEXT:    [[TMP145:%.*]] = load double*, double** [[C]], align 4
-// CHECK11-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP140]], i32* [[TMP147]], align 4
-// CHECK11-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP140]], i32* [[TMP149]], align 4
-// CHECK11-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP150]], align 4
-// CHECK11-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP142]], i32* [[TMP152]], align 4
-// CHECK11-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP142]], i32* [[TMP154]], align 4
-// CHECK11-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP155]], align 4
-// CHECK11-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to double**
-// CHECK11-NEXT:    store double* [[TMP143]], double** [[TMP157]], align 4
-// CHECK11-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to double**
-// CHECK11-NEXT:    store double* [[TMP143]], double** [[TMP159]], align 4
-// CHECK11-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP160]], align 4
-// CHECK11-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP162:%.*]] = bitcast i8** [[TMP161]] to double**
-// CHECK11-NEXT:    store double* [[TMP144]], double** [[TMP162]], align 4
-// CHECK11-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP164:%.*]] = bitcast i8** [[TMP163]] to double**
-// CHECK11-NEXT:    store double* [[TMP144]], double** [[TMP164]], align 4
-// CHECK11-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP165]], align 4
-// CHECK11-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to double**
-// CHECK11-NEXT:    store double* [[TMP145]], double** [[TMP167]], align 4
-// CHECK11-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to double**
-// CHECK11-NEXT:    store double* [[TMP145]], double** [[TMP169]], align 4
-// CHECK11-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 4
-// CHECK11-NEXT:    store i8* null, i8** [[TMP170]], align 4
-// CHECK11-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP70:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[ADD13:%.*]] = add nsw i32 [[TMP70]], 1
+// CHECK11-NEXT:    [[TMP71:%.*]] = zext i32 [[ADD13]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP71]])
+// CHECK11-NEXT:    [[KERNEL_ARGS14:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP72]], align 4
+// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP73]], align 4
+// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP66]], i8*** [[TMP74]], align 4
+// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP67]], i8*** [[TMP75]], align 4
+// CHECK11-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP76]], align 4
+// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP77]], align 4
+// CHECK11-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP78]], align 4
+// CHECK11-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP79]], align 4
+// CHECK11-NEXT:    [[TMP80:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l408.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]])
+// CHECK11-NEXT:    [[TMP81:%.*]] = icmp ne i32 [[TMP80]], 0
+// CHECK11-NEXT:    br i1 [[TMP81]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
+// CHECK11:       omp_offload.failed15:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l408(i32 [[TMP42]], double* [[TMP43]], double* [[TMP44]], double* [[TMP45]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
+// CHECK11:       omp_offload.cont16:
+// CHECK11-NEXT:    [[TMP82:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK11-NEXT:    store i32 [[TMP82]], i32* [[CH_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP83:%.*]] = load i32, i32* [[CH_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP84:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP84]], i32* [[N_CASTED17]], align 4
+// CHECK11-NEXT:    [[TMP85:%.*]] = load i32, i32* [[N_CASTED17]], align 4
+// CHECK11-NEXT:    [[TMP86:%.*]] = load double*, double** [[A]], align 4
+// CHECK11-NEXT:    [[TMP87:%.*]] = load double*, double** [[B]], align 4
+// CHECK11-NEXT:    [[TMP88:%.*]] = load double*, double** [[C]], align 4
+// CHECK11-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP83]], i32* [[TMP90]], align 4
+// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP83]], i32* [[TMP92]], align 4
+// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP93]], align 4
+// CHECK11-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP85]], i32* [[TMP95]], align 4
+// CHECK11-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP85]], i32* [[TMP97]], align 4
+// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP98]], align 4
+// CHECK11-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP100:%.*]] = bitcast i8** [[TMP99]] to double**
+// CHECK11-NEXT:    store double* [[TMP86]], double** [[TMP100]], align 4
+// CHECK11-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to double**
+// CHECK11-NEXT:    store double* [[TMP86]], double** [[TMP102]], align 4
+// CHECK11-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP103]], align 4
+// CHECK11-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to double**
+// CHECK11-NEXT:    store double* [[TMP87]], double** [[TMP105]], align 4
+// CHECK11-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to double**
+// CHECK11-NEXT:    store double* [[TMP87]], double** [[TMP107]], align 4
+// CHECK11-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP108]], align 4
+// CHECK11-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to double**
+// CHECK11-NEXT:    store double* [[TMP88]], double** [[TMP110]], align 4
+// CHECK11-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to double**
+// CHECK11-NEXT:    store double* [[TMP88]], double** [[TMP112]], align 4
+// CHECK11-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 4
+// CHECK11-NEXT:    store i8* null, i8** [[TMP113]], align 4
+// CHECK11-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP116:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP116]], i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK11-NEXT:    [[TMP117:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK11-NEXT:    [[SUB24:%.*]] = sub nsw i32 [[TMP117]], 0
+// CHECK11-NEXT:    [[DIV25:%.*]] = sdiv i32 [[SUB24]], 1
+// CHECK11-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[DIV25]], 1
+// CHECK11-NEXT:    store i32 [[SUB26]], i32* [[DOTCAPTURE_EXPR_23]], align 4
+// CHECK11-NEXT:    [[TMP118:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_23]], align 4
+// CHECK11-NEXT:    [[ADD27:%.*]] = add nsw i32 [[TMP118]], 1
+// CHECK11-NEXT:    [[TMP119:%.*]] = zext i32 [[ADD27]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP119]])
+// CHECK11-NEXT:    [[KERNEL_ARGS28:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP120]], align 4
+// CHECK11-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP121]], align 4
+// CHECK11-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP114]], i8*** [[TMP122]], align 4
+// CHECK11-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP115]], i8*** [[TMP123]], align 4
+// CHECK11-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP124]], align 4
+// CHECK11-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP125]], align 4
+// CHECK11-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP126]], align 4
+// CHECK11-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP127]], align 4
+// CHECK11-NEXT:    [[TMP128:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l447.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]])
+// CHECK11-NEXT:    [[TMP129:%.*]] = icmp ne i32 [[TMP128]], 0
+// CHECK11-NEXT:    br i1 [[TMP129]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
+// CHECK11:       omp_offload.failed29:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l447(i32 [[TMP83]], i32 [[TMP85]], double* [[TMP86]], double* [[TMP87]], double* [[TMP88]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
+// CHECK11:       omp_offload.cont30:
+// CHECK11-NEXT:    [[TMP130:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP130]], i32* [[N_CASTED31]], align 4
+// CHECK11-NEXT:    [[TMP131:%.*]] = load i32, i32* [[N_CASTED31]], align 4
+// CHECK11-NEXT:    [[TMP132:%.*]] = load double*, double** [[A]], align 4
+// CHECK11-NEXT:    [[TMP133:%.*]] = load double*, double** [[B]], align 4
+// CHECK11-NEXT:    [[TMP134:%.*]] = load double*, double** [[C]], align 4
+// CHECK11-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP131]], i32* [[TMP136]], align 4
+// CHECK11-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP131]], i32* [[TMP138]], align 4
+// CHECK11-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP139]], align 4
+// CHECK11-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to double**
+// CHECK11-NEXT:    store double* [[TMP132]], double** [[TMP141]], align 4
+// CHECK11-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to double**
+// CHECK11-NEXT:    store double* [[TMP132]], double** [[TMP143]], align 4
+// CHECK11-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP144]], align 4
+// CHECK11-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to double**
+// CHECK11-NEXT:    store double* [[TMP133]], double** [[TMP146]], align 4
+// CHECK11-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to double**
+// CHECK11-NEXT:    store double* [[TMP133]], double** [[TMP148]], align 4
+// CHECK11-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP149]], align 4
+// CHECK11-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to double**
+// CHECK11-NEXT:    store double* [[TMP134]], double** [[TMP151]], align 4
+// CHECK11-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to double**
+// CHECK11-NEXT:    store double* [[TMP134]], double** [[TMP153]], align 4
+// CHECK11-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP154]], align 4
+// CHECK11-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP157:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP157]], i32* [[DOTCAPTURE_EXPR_36]], align 4
+// CHECK11-NEXT:    [[TMP158:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_36]], align 4
+// CHECK11-NEXT:    [[SUB38:%.*]] = sub nsw i32 [[TMP158]], 0
+// CHECK11-NEXT:    [[DIV39:%.*]] = sdiv i32 [[SUB38]], 1
+// CHECK11-NEXT:    [[SUB40:%.*]] = sub nsw i32 [[DIV39]], 1
+// CHECK11-NEXT:    store i32 [[SUB40]], i32* [[DOTCAPTURE_EXPR_37]], align 4
+// CHECK11-NEXT:    [[TMP159:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_37]], align 4
+// CHECK11-NEXT:    [[ADD41:%.*]] = add nsw i32 [[TMP159]], 1
+// CHECK11-NEXT:    [[TMP160:%.*]] = zext i32 [[ADD41]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP160]])
+// CHECK11-NEXT:    [[KERNEL_ARGS42:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP161]], align 4
+// CHECK11-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP162]], align 4
+// CHECK11-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP155]], i8*** [[TMP163]], align 4
+// CHECK11-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP156]], i8*** [[TMP164]], align 4
+// CHECK11-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP165]], align 4
+// CHECK11-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP166]], align 4
+// CHECK11-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP167]], align 4
+// CHECK11-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP168]], align 4
+// CHECK11-NEXT:    [[TMP169:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l478.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]])
+// CHECK11-NEXT:    [[TMP170:%.*]] = icmp ne i32 [[TMP169]], 0
+// CHECK11-NEXT:    br i1 [[TMP170]], label [[OMP_OFFLOAD_FAILED43:%.*]], label [[OMP_OFFLOAD_CONT44:%.*]]
+// CHECK11:       omp_offload.failed43:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l478(i32 [[TMP131]], double* [[TMP132]], double* [[TMP133]], double* [[TMP134]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT44]]
+// CHECK11:       omp_offload.cont44:
+// CHECK11-NEXT:    [[TMP171:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK11-NEXT:    store i32 [[TMP171]], i32* [[CH_CASTED45]], align 4
+// CHECK11-NEXT:    [[TMP172:%.*]] = load i32, i32* [[CH_CASTED45]], align 4
 // CHECK11-NEXT:    [[TMP173:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP173]], i32* [[DOTCAPTURE_EXPR_48]], align 4
-// CHECK11-NEXT:    [[TMP174:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_48]], align 4
-// CHECK11-NEXT:    [[SUB50:%.*]] = sub nsw i32 [[TMP174]], 0
-// CHECK11-NEXT:    [[DIV51:%.*]] = sdiv i32 [[SUB50]], 1
-// CHECK11-NEXT:    [[SUB52:%.*]] = sub nsw i32 [[DIV51]], 1
-// CHECK11-NEXT:    store i32 [[SUB52]], i32* [[DOTCAPTURE_EXPR_49]], align 4
-// CHECK11-NEXT:    [[TMP175:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_49]], align 4
-// CHECK11-NEXT:    [[ADD53:%.*]] = add nsw i32 [[TMP175]], 1
-// CHECK11-NEXT:    [[TMP176:%.*]] = zext i32 [[ADD53]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP176]])
-// CHECK11-NEXT:    [[TMP177:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l506.region_id, i32 5, i8** [[TMP171]], i8** [[TMP172]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP178:%.*]] = icmp ne i32 [[TMP177]], 0
-// CHECK11-NEXT:    br i1 [[TMP178]], label [[OMP_OFFLOAD_FAILED54:%.*]], label [[OMP_OFFLOAD_CONT55:%.*]]
-// CHECK11:       omp_offload.failed54:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l506(i32 [[TMP140]], i32 [[TMP142]], double* [[TMP143]], double* [[TMP144]], double* [[TMP145]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT55]]
-// CHECK11:       omp_offload.cont55:
-// CHECK11-NEXT:    [[TMP179:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP179]], i32* [[N_CASTED56]], align 4
-// CHECK11-NEXT:    [[TMP180:%.*]] = load i32, i32* [[N_CASTED56]], align 4
-// CHECK11-NEXT:    [[TMP181:%.*]] = load double*, double** [[A]], align 4
-// CHECK11-NEXT:    [[TMP182:%.*]] = load double*, double** [[B]], align 4
-// CHECK11-NEXT:    [[TMP183:%.*]] = load double*, double** [[C]], align 4
-// CHECK11-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP185:%.*]] = bitcast i8** [[TMP184]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP180]], i32* [[TMP185]], align 4
-// CHECK11-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP187:%.*]] = bitcast i8** [[TMP186]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP180]], i32* [[TMP187]], align 4
-// CHECK11-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP188]], align 4
-// CHECK11-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP190:%.*]] = bitcast i8** [[TMP189]] to double**
-// CHECK11-NEXT:    store double* [[TMP181]], double** [[TMP190]], align 4
-// CHECK11-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP192:%.*]] = bitcast i8** [[TMP191]] to double**
-// CHECK11-NEXT:    store double* [[TMP181]], double** [[TMP192]], align 4
-// CHECK11-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP193]], align 4
-// CHECK11-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP195:%.*]] = bitcast i8** [[TMP194]] to double**
-// CHECK11-NEXT:    store double* [[TMP182]], double** [[TMP195]], align 4
-// CHECK11-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP197:%.*]] = bitcast i8** [[TMP196]] to double**
-// CHECK11-NEXT:    store double* [[TMP182]], double** [[TMP197]], align 4
-// CHECK11-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP198]], align 4
-// CHECK11-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP200:%.*]] = bitcast i8** [[TMP199]] to double**
-// CHECK11-NEXT:    store double* [[TMP183]], double** [[TMP200]], align 4
-// CHECK11-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP202:%.*]] = bitcast i8** [[TMP201]] to double**
-// CHECK11-NEXT:    store double* [[TMP183]], double** [[TMP202]], align 4
-// CHECK11-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP203]], align 4
-// CHECK11-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP206:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP206]], i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK11-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK11-NEXT:    [[SUB63:%.*]] = sub nsw i32 [[TMP207]], 0
-// CHECK11-NEXT:    [[DIV64:%.*]] = sdiv i32 [[SUB63]], 1
-// CHECK11-NEXT:    [[SUB65:%.*]] = sub nsw i32 [[DIV64]], 1
-// CHECK11-NEXT:    store i32 [[SUB65]], i32* [[DOTCAPTURE_EXPR_62]], align 4
-// CHECK11-NEXT:    [[TMP208:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_62]], align 4
-// CHECK11-NEXT:    [[ADD66:%.*]] = add nsw i32 [[TMP208]], 1
-// CHECK11-NEXT:    [[TMP209:%.*]] = zext i32 [[ADD66]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP209]])
-// CHECK11-NEXT:    [[TMP210:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l536.region_id, i32 4, i8** [[TMP204]], i8** [[TMP205]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP211:%.*]] = icmp ne i32 [[TMP210]], 0
-// CHECK11-NEXT:    br i1 [[TMP211]], label [[OMP_OFFLOAD_FAILED67:%.*]], label [[OMP_OFFLOAD_CONT68:%.*]]
-// CHECK11:       omp_offload.failed67:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l536(i32 [[TMP180]], double* [[TMP181]], double* [[TMP182]], double* [[TMP183]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT68]]
-// CHECK11:       omp_offload.cont68:
-// CHECK11-NEXT:    [[TMP212:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK11-NEXT:    store i32 [[TMP212]], i32* [[CH_CASTED69]], align 4
-// CHECK11-NEXT:    [[TMP213:%.*]] = load i32, i32* [[CH_CASTED69]], align 4
-// CHECK11-NEXT:    [[TMP214:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP214]], i32* [[N_CASTED70]], align 4
-// CHECK11-NEXT:    [[TMP215:%.*]] = load i32, i32* [[N_CASTED70]], align 4
-// CHECK11-NEXT:    [[TMP216:%.*]] = load double*, double** [[A]], align 4
-// CHECK11-NEXT:    [[TMP217:%.*]] = load double*, double** [[B]], align 4
-// CHECK11-NEXT:    [[TMP218:%.*]] = load double*, double** [[C]], align 4
-// CHECK11-NEXT:    [[TMP219:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP220:%.*]] = bitcast i8** [[TMP219]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP213]], i32* [[TMP220]], align 4
-// CHECK11-NEXT:    [[TMP221:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP222:%.*]] = bitcast i8** [[TMP221]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP213]], i32* [[TMP222]], align 4
-// CHECK11-NEXT:    [[TMP223:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP223]], align 4
-// CHECK11-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 [[TMP173]], i32* [[N_CASTED46]], align 4
+// CHECK11-NEXT:    [[TMP174:%.*]] = load i32, i32* [[N_CASTED46]], align 4
+// CHECK11-NEXT:    [[TMP175:%.*]] = load double*, double** [[A]], align 4
+// CHECK11-NEXT:    [[TMP176:%.*]] = load double*, double** [[B]], align 4
+// CHECK11-NEXT:    [[TMP177:%.*]] = load double*, double** [[C]], align 4
+// CHECK11-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP172]], i32* [[TMP179]], align 4
+// CHECK11-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP172]], i32* [[TMP181]], align 4
+// CHECK11-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP182]], align 4
+// CHECK11-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP184:%.*]] = bitcast i8** [[TMP183]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP174]], i32* [[TMP184]], align 4
+// CHECK11-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP186:%.*]] = bitcast i8** [[TMP185]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP174]], i32* [[TMP186]], align 4
+// CHECK11-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP187]], align 4
+// CHECK11-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP189:%.*]] = bitcast i8** [[TMP188]] to double**
+// CHECK11-NEXT:    store double* [[TMP175]], double** [[TMP189]], align 4
+// CHECK11-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP191:%.*]] = bitcast i8** [[TMP190]] to double**
+// CHECK11-NEXT:    store double* [[TMP175]], double** [[TMP191]], align 4
+// CHECK11-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP192]], align 4
+// CHECK11-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP194:%.*]] = bitcast i8** [[TMP193]] to double**
+// CHECK11-NEXT:    store double* [[TMP176]], double** [[TMP194]], align 4
+// CHECK11-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP196:%.*]] = bitcast i8** [[TMP195]] to double**
+// CHECK11-NEXT:    store double* [[TMP176]], double** [[TMP196]], align 4
+// CHECK11-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP197]], align 4
+// CHECK11-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP199:%.*]] = bitcast i8** [[TMP198]] to double**
+// CHECK11-NEXT:    store double* [[TMP177]], double** [[TMP199]], align 4
+// CHECK11-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP201:%.*]] = bitcast i8** [[TMP200]] to double**
+// CHECK11-NEXT:    store double* [[TMP177]], double** [[TMP201]], align 4
+// CHECK11-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 4
+// CHECK11-NEXT:    store i8* null, i8** [[TMP202]], align 4
+// CHECK11-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP205:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP205]], i32* [[DOTCAPTURE_EXPR_51]], align 4
+// CHECK11-NEXT:    [[TMP206:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_51]], align 4
+// CHECK11-NEXT:    [[SUB53:%.*]] = sub nsw i32 [[TMP206]], 0
+// CHECK11-NEXT:    [[DIV54:%.*]] = sdiv i32 [[SUB53]], 1
+// CHECK11-NEXT:    [[SUB55:%.*]] = sub nsw i32 [[DIV54]], 1
+// CHECK11-NEXT:    store i32 [[SUB55]], i32* [[DOTCAPTURE_EXPR_52]], align 4
+// CHECK11-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_52]], align 4
+// CHECK11-NEXT:    [[ADD56:%.*]] = add nsw i32 [[TMP207]], 1
+// CHECK11-NEXT:    [[TMP208:%.*]] = zext i32 [[ADD56]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP208]])
+// CHECK11-NEXT:    [[KERNEL_ARGS57:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP209:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP209]], align 4
+// CHECK11-NEXT:    [[TMP210:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP210]], align 4
+// CHECK11-NEXT:    [[TMP211:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP203]], i8*** [[TMP211]], align 4
+// CHECK11-NEXT:    [[TMP212:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP204]], i8*** [[TMP212]], align 4
+// CHECK11-NEXT:    [[TMP213:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP213]], align 4
+// CHECK11-NEXT:    [[TMP214:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP214]], align 4
+// CHECK11-NEXT:    [[TMP215:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP215]], align 4
+// CHECK11-NEXT:    [[TMP216:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP216]], align 4
+// CHECK11-NEXT:    [[TMP217:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l506.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]])
+// CHECK11-NEXT:    [[TMP218:%.*]] = icmp ne i32 [[TMP217]], 0
+// CHECK11-NEXT:    br i1 [[TMP218]], label [[OMP_OFFLOAD_FAILED58:%.*]], label [[OMP_OFFLOAD_CONT59:%.*]]
+// CHECK11:       omp_offload.failed58:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l506(i32 [[TMP172]], i32 [[TMP174]], double* [[TMP175]], double* [[TMP176]], double* [[TMP177]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT59]]
+// CHECK11:       omp_offload.cont59:
+// CHECK11-NEXT:    [[TMP219:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP219]], i32* [[N_CASTED60]], align 4
+// CHECK11-NEXT:    [[TMP220:%.*]] = load i32, i32* [[N_CASTED60]], align 4
+// CHECK11-NEXT:    [[TMP221:%.*]] = load double*, double** [[A]], align 4
+// CHECK11-NEXT:    [[TMP222:%.*]] = load double*, double** [[B]], align 4
+// CHECK11-NEXT:    [[TMP223:%.*]] = load double*, double** [[C]], align 4
+// CHECK11-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP225:%.*]] = bitcast i8** [[TMP224]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP215]], i32* [[TMP225]], align 4
-// CHECK11-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 [[TMP220]], i32* [[TMP225]], align 4
+// CHECK11-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP227:%.*]] = bitcast i8** [[TMP226]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP215]], i32* [[TMP227]], align 4
-// CHECK11-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 [[TMP220]], i32* [[TMP227]], align 4
+// CHECK11-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 0
 // CHECK11-NEXT:    store i8* null, i8** [[TMP228]], align 4
-// CHECK11-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 1
 // CHECK11-NEXT:    [[TMP230:%.*]] = bitcast i8** [[TMP229]] to double**
-// CHECK11-NEXT:    store double* [[TMP216]], double** [[TMP230]], align 4
-// CHECK11-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 2
+// CHECK11-NEXT:    store double* [[TMP221]], double** [[TMP230]], align 4
+// CHECK11-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 1
 // CHECK11-NEXT:    [[TMP232:%.*]] = bitcast i8** [[TMP231]] to double**
-// CHECK11-NEXT:    store double* [[TMP216]], double** [[TMP232]], align 4
-// CHECK11-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 2
+// CHECK11-NEXT:    store double* [[TMP221]], double** [[TMP232]], align 4
+// CHECK11-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 1
 // CHECK11-NEXT:    store i8* null, i8** [[TMP233]], align 4
-// CHECK11-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 2
 // CHECK11-NEXT:    [[TMP235:%.*]] = bitcast i8** [[TMP234]] to double**
-// CHECK11-NEXT:    store double* [[TMP217]], double** [[TMP235]], align 4
-// CHECK11-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 3
+// CHECK11-NEXT:    store double* [[TMP222]], double** [[TMP235]], align 4
+// CHECK11-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 2
 // CHECK11-NEXT:    [[TMP237:%.*]] = bitcast i8** [[TMP236]] to double**
-// CHECK11-NEXT:    store double* [[TMP217]], double** [[TMP237]], align 4
-// CHECK11-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 3
+// CHECK11-NEXT:    store double* [[TMP222]], double** [[TMP237]], align 4
+// CHECK11-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 2
 // CHECK11-NEXT:    store i8* null, i8** [[TMP238]], align 4
-// CHECK11-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 3
 // CHECK11-NEXT:    [[TMP240:%.*]] = bitcast i8** [[TMP239]] to double**
-// CHECK11-NEXT:    store double* [[TMP218]], double** [[TMP240]], align 4
-// CHECK11-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 4
+// CHECK11-NEXT:    store double* [[TMP223]], double** [[TMP240]], align 4
+// CHECK11-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 3
 // CHECK11-NEXT:    [[TMP242:%.*]] = bitcast i8** [[TMP241]] to double**
-// CHECK11-NEXT:    store double* [[TMP218]], double** [[TMP242]], align 4
-// CHECK11-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 4
+// CHECK11-NEXT:    store double* [[TMP223]], double** [[TMP242]], align 4
+// CHECK11-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 3
 // CHECK11-NEXT:    store i8* null, i8** [[TMP243]], align 4
-// CHECK11-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP246:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_75]], align 4
-// CHECK11-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_75]], align 4
-// CHECK11-NEXT:    [[SUB77:%.*]] = sub nsw i32 [[TMP247]], 0
-// CHECK11-NEXT:    [[DIV78:%.*]] = sdiv i32 [[SUB77]], 1
-// CHECK11-NEXT:    [[SUB79:%.*]] = sub nsw i32 [[DIV78]], 1
-// CHECK11-NEXT:    store i32 [[SUB79]], i32* [[DOTCAPTURE_EXPR_76]], align 4
-// CHECK11-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_76]], align 4
-// CHECK11-NEXT:    [[ADD80:%.*]] = add nsw i32 [[TMP248]], 1
-// CHECK11-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD80]] to i64
+// CHECK11-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_65]], align 4
+// CHECK11-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_65]], align 4
+// CHECK11-NEXT:    [[SUB67:%.*]] = sub nsw i32 [[TMP247]], 0
+// CHECK11-NEXT:    [[DIV68:%.*]] = sdiv i32 [[SUB67]], 1
+// CHECK11-NEXT:    [[SUB69:%.*]] = sub nsw i32 [[DIV68]], 1
+// CHECK11-NEXT:    store i32 [[SUB69]], i32* [[DOTCAPTURE_EXPR_66]], align 4
+// CHECK11-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_66]], align 4
+// CHECK11-NEXT:    [[ADD70:%.*]] = add nsw i32 [[TMP248]], 1
+// CHECK11-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD70]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP249]])
-// CHECK11-NEXT:    [[TMP250:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l562.region_id, i32 5, i8** [[TMP244]], i8** [[TMP245]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP251:%.*]] = icmp ne i32 [[TMP250]], 0
-// CHECK11-NEXT:    br i1 [[TMP251]], label [[OMP_OFFLOAD_FAILED81:%.*]], label [[OMP_OFFLOAD_CONT82:%.*]]
-// CHECK11:       omp_offload.failed81:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l562(i32 [[TMP213]], i32 [[TMP215]], double* [[TMP216]], double* [[TMP217]], double* [[TMP218]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT82]]
-// CHECK11:       omp_offload.cont82:
+// CHECK11-NEXT:    [[KERNEL_ARGS71:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP250:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP250]], align 4
+// CHECK11-NEXT:    [[TMP251:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP251]], align 4
+// CHECK11-NEXT:    [[TMP252:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP244]], i8*** [[TMP252]], align 4
+// CHECK11-NEXT:    [[TMP253:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP245]], i8*** [[TMP253]], align 4
+// CHECK11-NEXT:    [[TMP254:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP254]], align 4
+// CHECK11-NEXT:    [[TMP255:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP255]], align 4
+// CHECK11-NEXT:    [[TMP256:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP256]], align 4
+// CHECK11-NEXT:    [[TMP257:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP257]], align 4
+// CHECK11-NEXT:    [[TMP258:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l536.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]])
+// CHECK11-NEXT:    [[TMP259:%.*]] = icmp ne i32 [[TMP258]], 0
+// CHECK11-NEXT:    br i1 [[TMP259]], label [[OMP_OFFLOAD_FAILED72:%.*]], label [[OMP_OFFLOAD_CONT73:%.*]]
+// CHECK11:       omp_offload.failed72:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l536(i32 [[TMP220]], double* [[TMP221]], double* [[TMP222]], double* [[TMP223]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT73]]
+// CHECK11:       omp_offload.cont73:
+// CHECK11-NEXT:    [[TMP260:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK11-NEXT:    store i32 [[TMP260]], i32* [[CH_CASTED74]], align 4
+// CHECK11-NEXT:    [[TMP261:%.*]] = load i32, i32* [[CH_CASTED74]], align 4
+// CHECK11-NEXT:    [[TMP262:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP262]], i32* [[N_CASTED75]], align 4
+// CHECK11-NEXT:    [[TMP263:%.*]] = load i32, i32* [[N_CASTED75]], align 4
+// CHECK11-NEXT:    [[TMP264:%.*]] = load double*, double** [[A]], align 4
+// CHECK11-NEXT:    [[TMP265:%.*]] = load double*, double** [[B]], align 4
+// CHECK11-NEXT:    [[TMP266:%.*]] = load double*, double** [[C]], align 4
+// CHECK11-NEXT:    [[TMP267:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP268:%.*]] = bitcast i8** [[TMP267]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP261]], i32* [[TMP268]], align 4
+// CHECK11-NEXT:    [[TMP269:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP270:%.*]] = bitcast i8** [[TMP269]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP261]], i32* [[TMP270]], align 4
+// CHECK11-NEXT:    [[TMP271:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP271]], align 4
+// CHECK11-NEXT:    [[TMP272:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP273:%.*]] = bitcast i8** [[TMP272]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP263]], i32* [[TMP273]], align 4
+// CHECK11-NEXT:    [[TMP274:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP275:%.*]] = bitcast i8** [[TMP274]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP263]], i32* [[TMP275]], align 4
+// CHECK11-NEXT:    [[TMP276:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP276]], align 4
+// CHECK11-NEXT:    [[TMP277:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP278:%.*]] = bitcast i8** [[TMP277]] to double**
+// CHECK11-NEXT:    store double* [[TMP264]], double** [[TMP278]], align 4
+// CHECK11-NEXT:    [[TMP279:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP280:%.*]] = bitcast i8** [[TMP279]] to double**
+// CHECK11-NEXT:    store double* [[TMP264]], double** [[TMP280]], align 4
+// CHECK11-NEXT:    [[TMP281:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP281]], align 4
+// CHECK11-NEXT:    [[TMP282:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP283:%.*]] = bitcast i8** [[TMP282]] to double**
+// CHECK11-NEXT:    store double* [[TMP265]], double** [[TMP283]], align 4
+// CHECK11-NEXT:    [[TMP284:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP285:%.*]] = bitcast i8** [[TMP284]] to double**
+// CHECK11-NEXT:    store double* [[TMP265]], double** [[TMP285]], align 4
+// CHECK11-NEXT:    [[TMP286:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP286]], align 4
+// CHECK11-NEXT:    [[TMP287:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP288:%.*]] = bitcast i8** [[TMP287]] to double**
+// CHECK11-NEXT:    store double* [[TMP266]], double** [[TMP288]], align 4
+// CHECK11-NEXT:    [[TMP289:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP290:%.*]] = bitcast i8** [[TMP289]] to double**
+// CHECK11-NEXT:    store double* [[TMP266]], double** [[TMP290]], align 4
+// CHECK11-NEXT:    [[TMP291:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 4
+// CHECK11-NEXT:    store i8* null, i8** [[TMP291]], align 4
+// CHECK11-NEXT:    [[TMP292:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP293:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP294:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP294]], i32* [[DOTCAPTURE_EXPR_80]], align 4
+// CHECK11-NEXT:    [[TMP295:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_80]], align 4
+// CHECK11-NEXT:    [[SUB82:%.*]] = sub nsw i32 [[TMP295]], 0
+// CHECK11-NEXT:    [[DIV83:%.*]] = sdiv i32 [[SUB82]], 1
+// CHECK11-NEXT:    [[SUB84:%.*]] = sub nsw i32 [[DIV83]], 1
+// CHECK11-NEXT:    store i32 [[SUB84]], i32* [[DOTCAPTURE_EXPR_81]], align 4
+// CHECK11-NEXT:    [[TMP296:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_81]], align 4
+// CHECK11-NEXT:    [[ADD85:%.*]] = add nsw i32 [[TMP296]], 1
+// CHECK11-NEXT:    [[TMP297:%.*]] = zext i32 [[ADD85]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP297]])
+// CHECK11-NEXT:    [[KERNEL_ARGS86:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP298:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP298]], align 4
+// CHECK11-NEXT:    [[TMP299:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP299]], align 4
+// CHECK11-NEXT:    [[TMP300:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP292]], i8*** [[TMP300]], align 4
+// CHECK11-NEXT:    [[TMP301:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP293]], i8*** [[TMP301]], align 4
+// CHECK11-NEXT:    [[TMP302:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP302]], align 4
+// CHECK11-NEXT:    [[TMP303:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP303]], align 4
+// CHECK11-NEXT:    [[TMP304:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP304]], align 4
+// CHECK11-NEXT:    [[TMP305:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP305]], align 4
+// CHECK11-NEXT:    [[TMP306:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l562.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]])
+// CHECK11-NEXT:    [[TMP307:%.*]] = icmp ne i32 [[TMP306]], 0
+// CHECK11-NEXT:    br i1 [[TMP307]], label [[OMP_OFFLOAD_FAILED87:%.*]], label [[OMP_OFFLOAD_CONT88:%.*]]
+// CHECK11:       omp_offload.failed87:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l562(i32 [[TMP261]], i32 [[TMP263]], double* [[TMP264]], double* [[TMP265]], double* [[TMP266]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT88]]
+// CHECK11:       omp_offload.cont88:
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    ret i32 [[CALL]]
 //
@@ -10868,43 +11225,43 @@ int main() {
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_8:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[CH_CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED16:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_21:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED17:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS18:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS19:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS20:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP21:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED29:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS30:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS31:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS32:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP33:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_34:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_35:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[CH_CASTED42:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED43:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS44:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS45:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS46:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP47:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_48:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_49:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED56:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS57:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS58:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS59:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP60:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_61:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_62:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[CH_CASTED69:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED70:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS71:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS72:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS73:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP74:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_75:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_76:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_23:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED31:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS32:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS33:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS34:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP35:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_36:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_37:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[CH_CASTED45:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED46:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS47:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS48:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS49:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP50:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_51:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_52:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED60:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS61:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS62:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS63:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP64:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_65:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_66:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[CH_CASTED74:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED75:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS76:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS77:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS78:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP79:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_80:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_81:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 10000, i32* [[N]], align 4
 // CHECK11-NEXT:    store i32 100, i32* [[CH]], align 4
 // CHECK11-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N]], align 4
@@ -10958,394 +11315,513 @@ int main() {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP29]], 1
 // CHECK11-NEXT:    [[TMP30:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP30]])
-// CHECK11-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l42.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK11-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP25]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP26]], i8*** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l42.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK11-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l42(i32 [[TMP1]], i32* [[TMP2]], i32* [[TMP3]], i32* [[TMP4]]) #[[ATTR2]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP34:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = load i32*, i32** [[A]], align 4
-// CHECK11-NEXT:    [[TMP36:%.*]] = load i32*, i32** [[B]], align 4
-// CHECK11-NEXT:    [[TMP37:%.*]] = load i32*, i32** [[C]], align 4
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP34]], i32* [[TMP39]], align 4
-// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP34]], i32* [[TMP41]], align 4
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP42]], align 4
-// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP35]], i32** [[TMP44]], align 4
-// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP35]], i32** [[TMP46]], align 4
-// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP47]], align 4
-// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP36]], i32** [[TMP49]], align 4
-// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP36]], i32** [[TMP51]], align 4
-// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = load i32*, i32** [[A]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = load i32*, i32** [[B]], align 4
+// CHECK11-NEXT:    [[TMP45:%.*]] = load i32*, i32** [[C]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP42]], i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP42]], i32* [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP43]], i32** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
 // CHECK11-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP37]], i32** [[TMP54]], align 4
-// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP37]], i32** [[TMP56]], align 4
-// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP57]], align 4
-// CHECK11-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP60:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP60]], i32* [[DOTCAPTURE_EXPR_8]], align 4
-// CHECK11-NEXT:    [[TMP61:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_8]], align 4
-// CHECK11-NEXT:    [[SUB10:%.*]] = sub nsw i32 [[TMP61]], 0
+// CHECK11-NEXT:    store i32* [[TMP43]], i32** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP55]], align 4
+// CHECK11-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP44]], i32** [[TMP57]], align 4
+// CHECK11-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP44]], i32** [[TMP59]], align 4
+// CHECK11-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP45]], i32** [[TMP62]], align 4
+// CHECK11-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP64:%.*]] = bitcast i8** [[TMP63]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP45]], i32** [[TMP64]], align 4
+// CHECK11-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP65]], align 4
+// CHECK11-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP68]], i32* [[DOTCAPTURE_EXPR_8]], align 4
+// CHECK11-NEXT:    [[TMP69:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_8]], align 4
+// CHECK11-NEXT:    [[SUB10:%.*]] = sub nsw i32 [[TMP69]], 0
 // CHECK11-NEXT:    [[DIV11:%.*]] = sdiv i32 [[SUB10]], 1
 // CHECK11-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[DIV11]], 1
 // CHECK11-NEXT:    store i32 [[SUB12]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[TMP62:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[ADD13:%.*]] = add nsw i32 [[TMP62]], 1
-// CHECK11-NEXT:    [[TMP63:%.*]] = zext i32 [[ADD13]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP63]])
-// CHECK11-NEXT:    [[TMP64:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l51.region_id, i32 4, i8** [[TMP58]], i8** [[TMP59]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP65:%.*]] = icmp ne i32 [[TMP64]], 0
-// CHECK11-NEXT:    br i1 [[TMP65]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK11:       omp_offload.failed14:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l51(i32 [[TMP34]], i32* [[TMP35]], i32* [[TMP36]], i32* [[TMP37]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK11:       omp_offload.cont15:
-// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK11-NEXT:    store i32 [[TMP66]], i32* [[CH_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP67:%.*]] = load i32, i32* [[CH_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP68]], i32* [[N_CASTED16]], align 4
-// CHECK11-NEXT:    [[TMP69:%.*]] = load i32, i32* [[N_CASTED16]], align 4
-// CHECK11-NEXT:    [[TMP70:%.*]] = load i32*, i32** [[A]], align 4
-// CHECK11-NEXT:    [[TMP71:%.*]] = load i32*, i32** [[B]], align 4
-// CHECK11-NEXT:    [[TMP72:%.*]] = load i32*, i32** [[C]], align 4
-// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP67]], i32* [[TMP74]], align 4
-// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP67]], i32* [[TMP76]], align 4
-// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP77]], align 4
-// CHECK11-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP69]], i32* [[TMP79]], align 4
-// CHECK11-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP69]], i32* [[TMP81]], align 4
-// CHECK11-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP82]], align 4
-// CHECK11-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP70]], i32** [[TMP84]], align 4
-// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP70]], i32** [[TMP86]], align 4
-// CHECK11-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP87]], align 4
-// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP71]], i32** [[TMP89]], align 4
-// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP71]], i32** [[TMP91]], align 4
-// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP92]], align 4
-// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP72]], i32** [[TMP94]], align 4
-// CHECK11-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP72]], i32** [[TMP96]], align 4
-// CHECK11-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 4
-// CHECK11-NEXT:    store i8* null, i8** [[TMP97]], align 4
-// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP100]], i32* [[DOTCAPTURE_EXPR_21]], align 4
-// CHECK11-NEXT:    [[TMP101:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_21]], align 4
-// CHECK11-NEXT:    [[SUB23:%.*]] = sub nsw i32 [[TMP101]], 0
-// CHECK11-NEXT:    [[DIV24:%.*]] = sdiv i32 [[SUB23]], 1
-// CHECK11-NEXT:    [[SUB25:%.*]] = sub nsw i32 [[DIV24]], 1
-// CHECK11-NEXT:    store i32 [[SUB25]], i32* [[DOTCAPTURE_EXPR_22]], align 4
-// CHECK11-NEXT:    [[TMP102:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
-// CHECK11-NEXT:    [[ADD26:%.*]] = add nsw i32 [[TMP102]], 1
-// CHECK11-NEXT:    [[TMP103:%.*]] = zext i32 [[ADD26]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP103]])
-// CHECK11-NEXT:    [[TMP104:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l59.region_id, i32 5, i8** [[TMP98]], i8** [[TMP99]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP105:%.*]] = icmp ne i32 [[TMP104]], 0
-// CHECK11-NEXT:    br i1 [[TMP105]], label [[OMP_OFFLOAD_FAILED27:%.*]], label [[OMP_OFFLOAD_CONT28:%.*]]
-// CHECK11:       omp_offload.failed27:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l59(i32 [[TMP67]], i32 [[TMP69]], i32* [[TMP70]], i32* [[TMP71]], i32* [[TMP72]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT28]]
-// CHECK11:       omp_offload.cont28:
-// CHECK11-NEXT:    [[TMP106:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP106]], i32* [[N_CASTED29]], align 4
-// CHECK11-NEXT:    [[TMP107:%.*]] = load i32, i32* [[N_CASTED29]], align 4
-// CHECK11-NEXT:    [[TMP108:%.*]] = load i32*, i32** [[A]], align 4
-// CHECK11-NEXT:    [[TMP109:%.*]] = load i32*, i32** [[B]], align 4
-// CHECK11-NEXT:    [[TMP110:%.*]] = load i32*, i32** [[C]], align 4
-// CHECK11-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP107]], i32* [[TMP112]], align 4
-// CHECK11-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP107]], i32* [[TMP114]], align 4
-// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP115]], align 4
-// CHECK11-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP108]], i32** [[TMP117]], align 4
-// CHECK11-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP119:%.*]] = bitcast i8** [[TMP118]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP108]], i32** [[TMP119]], align 4
-// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP120]], align 4
-// CHECK11-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP109]], i32** [[TMP122]], align 4
-// CHECK11-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP109]], i32** [[TMP124]], align 4
-// CHECK11-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP125]], align 4
-// CHECK11-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP110]], i32** [[TMP127]], align 4
-// CHECK11-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP110]], i32** [[TMP129]], align 4
-// CHECK11-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP130]], align 4
-// CHECK11-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP133:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP133]], i32* [[DOTCAPTURE_EXPR_34]], align 4
-// CHECK11-NEXT:    [[TMP134:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_34]], align 4
-// CHECK11-NEXT:    [[SUB36:%.*]] = sub nsw i32 [[TMP134]], 0
-// CHECK11-NEXT:    [[DIV37:%.*]] = sdiv i32 [[SUB36]], 1
-// CHECK11-NEXT:    [[SUB38:%.*]] = sub nsw i32 [[DIV37]], 1
-// CHECK11-NEXT:    store i32 [[SUB38]], i32* [[DOTCAPTURE_EXPR_35]], align 4
-// CHECK11-NEXT:    [[TMP135:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_35]], align 4
-// CHECK11-NEXT:    [[ADD39:%.*]] = add nsw i32 [[TMP135]], 1
-// CHECK11-NEXT:    [[TMP136:%.*]] = zext i32 [[ADD39]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP136]])
-// CHECK11-NEXT:    [[TMP137:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l67.region_id, i32 4, i8** [[TMP131]], i8** [[TMP132]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.40, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.41, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP138:%.*]] = icmp ne i32 [[TMP137]], 0
-// CHECK11-NEXT:    br i1 [[TMP138]], label [[OMP_OFFLOAD_FAILED40:%.*]], label [[OMP_OFFLOAD_CONT41:%.*]]
-// CHECK11:       omp_offload.failed40:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l67(i32 [[TMP107]], i32* [[TMP108]], i32* [[TMP109]], i32* [[TMP110]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT41]]
-// CHECK11:       omp_offload.cont41:
-// CHECK11-NEXT:    [[TMP139:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK11-NEXT:    store i32 [[TMP139]], i32* [[CH_CASTED42]], align 4
-// CHECK11-NEXT:    [[TMP140:%.*]] = load i32, i32* [[CH_CASTED42]], align 4
-// CHECK11-NEXT:    [[TMP141:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP141]], i32* [[N_CASTED43]], align 4
-// CHECK11-NEXT:    [[TMP142:%.*]] = load i32, i32* [[N_CASTED43]], align 4
-// CHECK11-NEXT:    [[TMP143:%.*]] = load i32*, i32** [[A]], align 4
-// CHECK11-NEXT:    [[TMP144:%.*]] = load i32*, i32** [[B]], align 4
-// CHECK11-NEXT:    [[TMP145:%.*]] = load i32*, i32** [[C]], align 4
-// CHECK11-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP140]], i32* [[TMP147]], align 4
-// CHECK11-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP140]], i32* [[TMP149]], align 4
-// CHECK11-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP150]], align 4
-// CHECK11-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP142]], i32* [[TMP152]], align 4
-// CHECK11-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP142]], i32* [[TMP154]], align 4
-// CHECK11-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP155]], align 4
-// CHECK11-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP143]], i32** [[TMP157]], align 4
-// CHECK11-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP143]], i32** [[TMP159]], align 4
-// CHECK11-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP160]], align 4
-// CHECK11-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP162:%.*]] = bitcast i8** [[TMP161]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP144]], i32** [[TMP162]], align 4
-// CHECK11-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP164:%.*]] = bitcast i8** [[TMP163]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP144]], i32** [[TMP164]], align 4
-// CHECK11-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP165]], align 4
-// CHECK11-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP145]], i32** [[TMP167]], align 4
-// CHECK11-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP145]], i32** [[TMP169]], align 4
-// CHECK11-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 4
-// CHECK11-NEXT:    store i8* null, i8** [[TMP170]], align 4
-// CHECK11-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP70:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[ADD13:%.*]] = add nsw i32 [[TMP70]], 1
+// CHECK11-NEXT:    [[TMP71:%.*]] = zext i32 [[ADD13]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP71]])
+// CHECK11-NEXT:    [[KERNEL_ARGS14:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP72]], align 4
+// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP73]], align 4
+// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP66]], i8*** [[TMP74]], align 4
+// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP67]], i8*** [[TMP75]], align 4
+// CHECK11-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP76]], align 4
+// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP77]], align 4
+// CHECK11-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP78]], align 4
+// CHECK11-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP79]], align 4
+// CHECK11-NEXT:    [[TMP80:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l51.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]])
+// CHECK11-NEXT:    [[TMP81:%.*]] = icmp ne i32 [[TMP80]], 0
+// CHECK11-NEXT:    br i1 [[TMP81]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
+// CHECK11:       omp_offload.failed15:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l51(i32 [[TMP42]], i32* [[TMP43]], i32* [[TMP44]], i32* [[TMP45]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
+// CHECK11:       omp_offload.cont16:
+// CHECK11-NEXT:    [[TMP82:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK11-NEXT:    store i32 [[TMP82]], i32* [[CH_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP83:%.*]] = load i32, i32* [[CH_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP84:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP84]], i32* [[N_CASTED17]], align 4
+// CHECK11-NEXT:    [[TMP85:%.*]] = load i32, i32* [[N_CASTED17]], align 4
+// CHECK11-NEXT:    [[TMP86:%.*]] = load i32*, i32** [[A]], align 4
+// CHECK11-NEXT:    [[TMP87:%.*]] = load i32*, i32** [[B]], align 4
+// CHECK11-NEXT:    [[TMP88:%.*]] = load i32*, i32** [[C]], align 4
+// CHECK11-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP83]], i32* [[TMP90]], align 4
+// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP83]], i32* [[TMP92]], align 4
+// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP93]], align 4
+// CHECK11-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP85]], i32* [[TMP95]], align 4
+// CHECK11-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP85]], i32* [[TMP97]], align 4
+// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP98]], align 4
+// CHECK11-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP100:%.*]] = bitcast i8** [[TMP99]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP86]], i32** [[TMP100]], align 4
+// CHECK11-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP86]], i32** [[TMP102]], align 4
+// CHECK11-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP103]], align 4
+// CHECK11-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP87]], i32** [[TMP105]], align 4
+// CHECK11-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP87]], i32** [[TMP107]], align 4
+// CHECK11-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP108]], align 4
+// CHECK11-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP88]], i32** [[TMP110]], align 4
+// CHECK11-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP88]], i32** [[TMP112]], align 4
+// CHECK11-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 4
+// CHECK11-NEXT:    store i8* null, i8** [[TMP113]], align 4
+// CHECK11-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP116:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP116]], i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK11-NEXT:    [[TMP117:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK11-NEXT:    [[SUB24:%.*]] = sub nsw i32 [[TMP117]], 0
+// CHECK11-NEXT:    [[DIV25:%.*]] = sdiv i32 [[SUB24]], 1
+// CHECK11-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[DIV25]], 1
+// CHECK11-NEXT:    store i32 [[SUB26]], i32* [[DOTCAPTURE_EXPR_23]], align 4
+// CHECK11-NEXT:    [[TMP118:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_23]], align 4
+// CHECK11-NEXT:    [[ADD27:%.*]] = add nsw i32 [[TMP118]], 1
+// CHECK11-NEXT:    [[TMP119:%.*]] = zext i32 [[ADD27]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP119]])
+// CHECK11-NEXT:    [[KERNEL_ARGS28:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP120]], align 4
+// CHECK11-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP121]], align 4
+// CHECK11-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP114]], i8*** [[TMP122]], align 4
+// CHECK11-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP115]], i8*** [[TMP123]], align 4
+// CHECK11-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP124]], align 4
+// CHECK11-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP125]], align 4
+// CHECK11-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP126]], align 4
+// CHECK11-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP127]], align 4
+// CHECK11-NEXT:    [[TMP128:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l59.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]])
+// CHECK11-NEXT:    [[TMP129:%.*]] = icmp ne i32 [[TMP128]], 0
+// CHECK11-NEXT:    br i1 [[TMP129]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
+// CHECK11:       omp_offload.failed29:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l59(i32 [[TMP83]], i32 [[TMP85]], i32* [[TMP86]], i32* [[TMP87]], i32* [[TMP88]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
+// CHECK11:       omp_offload.cont30:
+// CHECK11-NEXT:    [[TMP130:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP130]], i32* [[N_CASTED31]], align 4
+// CHECK11-NEXT:    [[TMP131:%.*]] = load i32, i32* [[N_CASTED31]], align 4
+// CHECK11-NEXT:    [[TMP132:%.*]] = load i32*, i32** [[A]], align 4
+// CHECK11-NEXT:    [[TMP133:%.*]] = load i32*, i32** [[B]], align 4
+// CHECK11-NEXT:    [[TMP134:%.*]] = load i32*, i32** [[C]], align 4
+// CHECK11-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP131]], i32* [[TMP136]], align 4
+// CHECK11-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP131]], i32* [[TMP138]], align 4
+// CHECK11-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP139]], align 4
+// CHECK11-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP132]], i32** [[TMP141]], align 4
+// CHECK11-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP132]], i32** [[TMP143]], align 4
+// CHECK11-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP144]], align 4
+// CHECK11-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP133]], i32** [[TMP146]], align 4
+// CHECK11-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP133]], i32** [[TMP148]], align 4
+// CHECK11-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP149]], align 4
+// CHECK11-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP134]], i32** [[TMP151]], align 4
+// CHECK11-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP134]], i32** [[TMP153]], align 4
+// CHECK11-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP154]], align 4
+// CHECK11-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP157:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP157]], i32* [[DOTCAPTURE_EXPR_36]], align 4
+// CHECK11-NEXT:    [[TMP158:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_36]], align 4
+// CHECK11-NEXT:    [[SUB38:%.*]] = sub nsw i32 [[TMP158]], 0
+// CHECK11-NEXT:    [[DIV39:%.*]] = sdiv i32 [[SUB38]], 1
+// CHECK11-NEXT:    [[SUB40:%.*]] = sub nsw i32 [[DIV39]], 1
+// CHECK11-NEXT:    store i32 [[SUB40]], i32* [[DOTCAPTURE_EXPR_37]], align 4
+// CHECK11-NEXT:    [[TMP159:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_37]], align 4
+// CHECK11-NEXT:    [[ADD41:%.*]] = add nsw i32 [[TMP159]], 1
+// CHECK11-NEXT:    [[TMP160:%.*]] = zext i32 [[ADD41]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP160]])
+// CHECK11-NEXT:    [[KERNEL_ARGS42:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP161]], align 4
+// CHECK11-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP162]], align 4
+// CHECK11-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP155]], i8*** [[TMP163]], align 4
+// CHECK11-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP156]], i8*** [[TMP164]], align 4
+// CHECK11-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.40, i32 0, i32 0), i64** [[TMP165]], align 4
+// CHECK11-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.41, i32 0, i32 0), i64** [[TMP166]], align 4
+// CHECK11-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP167]], align 4
+// CHECK11-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP168]], align 4
+// CHECK11-NEXT:    [[TMP169:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l67.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]])
+// CHECK11-NEXT:    [[TMP170:%.*]] = icmp ne i32 [[TMP169]], 0
+// CHECK11-NEXT:    br i1 [[TMP170]], label [[OMP_OFFLOAD_FAILED43:%.*]], label [[OMP_OFFLOAD_CONT44:%.*]]
+// CHECK11:       omp_offload.failed43:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l67(i32 [[TMP131]], i32* [[TMP132]], i32* [[TMP133]], i32* [[TMP134]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT44]]
+// CHECK11:       omp_offload.cont44:
+// CHECK11-NEXT:    [[TMP171:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK11-NEXT:    store i32 [[TMP171]], i32* [[CH_CASTED45]], align 4
+// CHECK11-NEXT:    [[TMP172:%.*]] = load i32, i32* [[CH_CASTED45]], align 4
 // CHECK11-NEXT:    [[TMP173:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP173]], i32* [[DOTCAPTURE_EXPR_48]], align 4
-// CHECK11-NEXT:    [[TMP174:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_48]], align 4
-// CHECK11-NEXT:    [[SUB50:%.*]] = sub nsw i32 [[TMP174]], 0
-// CHECK11-NEXT:    [[DIV51:%.*]] = sdiv i32 [[SUB50]], 1
-// CHECK11-NEXT:    [[SUB52:%.*]] = sub nsw i32 [[DIV51]], 1
-// CHECK11-NEXT:    store i32 [[SUB52]], i32* [[DOTCAPTURE_EXPR_49]], align 4
-// CHECK11-NEXT:    [[TMP175:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_49]], align 4
-// CHECK11-NEXT:    [[ADD53:%.*]] = add nsw i32 [[TMP175]], 1
-// CHECK11-NEXT:    [[TMP176:%.*]] = zext i32 [[ADD53]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP176]])
-// CHECK11-NEXT:    [[TMP177:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75.region_id, i32 5, i8** [[TMP171]], i8** [[TMP172]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.44, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.45, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP178:%.*]] = icmp ne i32 [[TMP177]], 0
-// CHECK11-NEXT:    br i1 [[TMP178]], label [[OMP_OFFLOAD_FAILED54:%.*]], label [[OMP_OFFLOAD_CONT55:%.*]]
-// CHECK11:       omp_offload.failed54:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75(i32 [[TMP140]], i32 [[TMP142]], i32* [[TMP143]], i32* [[TMP144]], i32* [[TMP145]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT55]]
-// CHECK11:       omp_offload.cont55:
-// CHECK11-NEXT:    [[TMP179:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP179]], i32* [[N_CASTED56]], align 4
-// CHECK11-NEXT:    [[TMP180:%.*]] = load i32, i32* [[N_CASTED56]], align 4
-// CHECK11-NEXT:    [[TMP181:%.*]] = load i32*, i32** [[A]], align 4
-// CHECK11-NEXT:    [[TMP182:%.*]] = load i32*, i32** [[B]], align 4
-// CHECK11-NEXT:    [[TMP183:%.*]] = load i32*, i32** [[C]], align 4
-// CHECK11-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP185:%.*]] = bitcast i8** [[TMP184]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP180]], i32* [[TMP185]], align 4
-// CHECK11-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP187:%.*]] = bitcast i8** [[TMP186]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP180]], i32* [[TMP187]], align 4
-// CHECK11-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP188]], align 4
-// CHECK11-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP190:%.*]] = bitcast i8** [[TMP189]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP181]], i32** [[TMP190]], align 4
-// CHECK11-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP192:%.*]] = bitcast i8** [[TMP191]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP181]], i32** [[TMP192]], align 4
-// CHECK11-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP193]], align 4
-// CHECK11-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP195:%.*]] = bitcast i8** [[TMP194]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP182]], i32** [[TMP195]], align 4
-// CHECK11-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP197:%.*]] = bitcast i8** [[TMP196]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP182]], i32** [[TMP197]], align 4
-// CHECK11-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP198]], align 4
-// CHECK11-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP200:%.*]] = bitcast i8** [[TMP199]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP183]], i32** [[TMP200]], align 4
-// CHECK11-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP202:%.*]] = bitcast i8** [[TMP201]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP183]], i32** [[TMP202]], align 4
-// CHECK11-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP203]], align 4
-// CHECK11-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP206:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP206]], i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK11-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK11-NEXT:    [[SUB63:%.*]] = sub nsw i32 [[TMP207]], 0
-// CHECK11-NEXT:    [[DIV64:%.*]] = sdiv i32 [[SUB63]], 1
-// CHECK11-NEXT:    [[SUB65:%.*]] = sub nsw i32 [[DIV64]], 1
-// CHECK11-NEXT:    store i32 [[SUB65]], i32* [[DOTCAPTURE_EXPR_62]], align 4
-// CHECK11-NEXT:    [[TMP208:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_62]], align 4
-// CHECK11-NEXT:    [[ADD66:%.*]] = add nsw i32 [[TMP208]], 1
-// CHECK11-NEXT:    [[TMP209:%.*]] = zext i32 [[ADD66]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP209]])
-// CHECK11-NEXT:    [[TMP210:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l83.region_id, i32 4, i8** [[TMP204]], i8** [[TMP205]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.48, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.49, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP211:%.*]] = icmp ne i32 [[TMP210]], 0
-// CHECK11-NEXT:    br i1 [[TMP211]], label [[OMP_OFFLOAD_FAILED67:%.*]], label [[OMP_OFFLOAD_CONT68:%.*]]
-// CHECK11:       omp_offload.failed67:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l83(i32 [[TMP180]], i32* [[TMP181]], i32* [[TMP182]], i32* [[TMP183]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT68]]
-// CHECK11:       omp_offload.cont68:
-// CHECK11-NEXT:    [[TMP212:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK11-NEXT:    store i32 [[TMP212]], i32* [[CH_CASTED69]], align 4
-// CHECK11-NEXT:    [[TMP213:%.*]] = load i32, i32* [[CH_CASTED69]], align 4
-// CHECK11-NEXT:    [[TMP214:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP214]], i32* [[N_CASTED70]], align 4
-// CHECK11-NEXT:    [[TMP215:%.*]] = load i32, i32* [[N_CASTED70]], align 4
-// CHECK11-NEXT:    [[TMP216:%.*]] = load i32*, i32** [[A]], align 4
-// CHECK11-NEXT:    [[TMP217:%.*]] = load i32*, i32** [[B]], align 4
-// CHECK11-NEXT:    [[TMP218:%.*]] = load i32*, i32** [[C]], align 4
-// CHECK11-NEXT:    [[TMP219:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP220:%.*]] = bitcast i8** [[TMP219]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP213]], i32* [[TMP220]], align 4
-// CHECK11-NEXT:    [[TMP221:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP222:%.*]] = bitcast i8** [[TMP221]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP213]], i32* [[TMP222]], align 4
-// CHECK11-NEXT:    [[TMP223:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP223]], align 4
-// CHECK11-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 [[TMP173]], i32* [[N_CASTED46]], align 4
+// CHECK11-NEXT:    [[TMP174:%.*]] = load i32, i32* [[N_CASTED46]], align 4
+// CHECK11-NEXT:    [[TMP175:%.*]] = load i32*, i32** [[A]], align 4
+// CHECK11-NEXT:    [[TMP176:%.*]] = load i32*, i32** [[B]], align 4
+// CHECK11-NEXT:    [[TMP177:%.*]] = load i32*, i32** [[C]], align 4
+// CHECK11-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP172]], i32* [[TMP179]], align 4
+// CHECK11-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP172]], i32* [[TMP181]], align 4
+// CHECK11-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP182]], align 4
+// CHECK11-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP184:%.*]] = bitcast i8** [[TMP183]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP174]], i32* [[TMP184]], align 4
+// CHECK11-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP186:%.*]] = bitcast i8** [[TMP185]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP174]], i32* [[TMP186]], align 4
+// CHECK11-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP187]], align 4
+// CHECK11-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP189:%.*]] = bitcast i8** [[TMP188]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP175]], i32** [[TMP189]], align 4
+// CHECK11-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP191:%.*]] = bitcast i8** [[TMP190]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP175]], i32** [[TMP191]], align 4
+// CHECK11-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP192]], align 4
+// CHECK11-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP194:%.*]] = bitcast i8** [[TMP193]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP176]], i32** [[TMP194]], align 4
+// CHECK11-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP196:%.*]] = bitcast i8** [[TMP195]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP176]], i32** [[TMP196]], align 4
+// CHECK11-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP197]], align 4
+// CHECK11-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP199:%.*]] = bitcast i8** [[TMP198]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP177]], i32** [[TMP199]], align 4
+// CHECK11-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP201:%.*]] = bitcast i8** [[TMP200]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP177]], i32** [[TMP201]], align 4
+// CHECK11-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 4
+// CHECK11-NEXT:    store i8* null, i8** [[TMP202]], align 4
+// CHECK11-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP205:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP205]], i32* [[DOTCAPTURE_EXPR_51]], align 4
+// CHECK11-NEXT:    [[TMP206:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_51]], align 4
+// CHECK11-NEXT:    [[SUB53:%.*]] = sub nsw i32 [[TMP206]], 0
+// CHECK11-NEXT:    [[DIV54:%.*]] = sdiv i32 [[SUB53]], 1
+// CHECK11-NEXT:    [[SUB55:%.*]] = sub nsw i32 [[DIV54]], 1
+// CHECK11-NEXT:    store i32 [[SUB55]], i32* [[DOTCAPTURE_EXPR_52]], align 4
+// CHECK11-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_52]], align 4
+// CHECK11-NEXT:    [[ADD56:%.*]] = add nsw i32 [[TMP207]], 1
+// CHECK11-NEXT:    [[TMP208:%.*]] = zext i32 [[ADD56]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP208]])
+// CHECK11-NEXT:    [[KERNEL_ARGS57:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP209:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP209]], align 4
+// CHECK11-NEXT:    [[TMP210:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP210]], align 4
+// CHECK11-NEXT:    [[TMP211:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP203]], i8*** [[TMP211]], align 4
+// CHECK11-NEXT:    [[TMP212:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP204]], i8*** [[TMP212]], align 4
+// CHECK11-NEXT:    [[TMP213:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.44, i32 0, i32 0), i64** [[TMP213]], align 4
+// CHECK11-NEXT:    [[TMP214:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.45, i32 0, i32 0), i64** [[TMP214]], align 4
+// CHECK11-NEXT:    [[TMP215:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP215]], align 4
+// CHECK11-NEXT:    [[TMP216:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP216]], align 4
+// CHECK11-NEXT:    [[TMP217:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]])
+// CHECK11-NEXT:    [[TMP218:%.*]] = icmp ne i32 [[TMP217]], 0
+// CHECK11-NEXT:    br i1 [[TMP218]], label [[OMP_OFFLOAD_FAILED58:%.*]], label [[OMP_OFFLOAD_CONT59:%.*]]
+// CHECK11:       omp_offload.failed58:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75(i32 [[TMP172]], i32 [[TMP174]], i32* [[TMP175]], i32* [[TMP176]], i32* [[TMP177]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT59]]
+// CHECK11:       omp_offload.cont59:
+// CHECK11-NEXT:    [[TMP219:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP219]], i32* [[N_CASTED60]], align 4
+// CHECK11-NEXT:    [[TMP220:%.*]] = load i32, i32* [[N_CASTED60]], align 4
+// CHECK11-NEXT:    [[TMP221:%.*]] = load i32*, i32** [[A]], align 4
+// CHECK11-NEXT:    [[TMP222:%.*]] = load i32*, i32** [[B]], align 4
+// CHECK11-NEXT:    [[TMP223:%.*]] = load i32*, i32** [[C]], align 4
+// CHECK11-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP225:%.*]] = bitcast i8** [[TMP224]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP215]], i32* [[TMP225]], align 4
-// CHECK11-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 [[TMP220]], i32* [[TMP225]], align 4
+// CHECK11-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP227:%.*]] = bitcast i8** [[TMP226]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP215]], i32* [[TMP227]], align 4
-// CHECK11-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 [[TMP220]], i32* [[TMP227]], align 4
+// CHECK11-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 0
 // CHECK11-NEXT:    store i8* null, i8** [[TMP228]], align 4
-// CHECK11-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 1
 // CHECK11-NEXT:    [[TMP230:%.*]] = bitcast i8** [[TMP229]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP216]], i32** [[TMP230]], align 4
-// CHECK11-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 2
+// CHECK11-NEXT:    store i32* [[TMP221]], i32** [[TMP230]], align 4
+// CHECK11-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 1
 // CHECK11-NEXT:    [[TMP232:%.*]] = bitcast i8** [[TMP231]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP216]], i32** [[TMP232]], align 4
-// CHECK11-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 2
+// CHECK11-NEXT:    store i32* [[TMP221]], i32** [[TMP232]], align 4
+// CHECK11-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 1
 // CHECK11-NEXT:    store i8* null, i8** [[TMP233]], align 4
-// CHECK11-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 2
 // CHECK11-NEXT:    [[TMP235:%.*]] = bitcast i8** [[TMP234]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP217]], i32** [[TMP235]], align 4
-// CHECK11-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 3
+// CHECK11-NEXT:    store i32* [[TMP222]], i32** [[TMP235]], align 4
+// CHECK11-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 2
 // CHECK11-NEXT:    [[TMP237:%.*]] = bitcast i8** [[TMP236]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP217]], i32** [[TMP237]], align 4
-// CHECK11-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 3
+// CHECK11-NEXT:    store i32* [[TMP222]], i32** [[TMP237]], align 4
+// CHECK11-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 2
 // CHECK11-NEXT:    store i8* null, i8** [[TMP238]], align 4
-// CHECK11-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 3
 // CHECK11-NEXT:    [[TMP240:%.*]] = bitcast i8** [[TMP239]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP218]], i32** [[TMP240]], align 4
-// CHECK11-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 4
+// CHECK11-NEXT:    store i32* [[TMP223]], i32** [[TMP240]], align 4
+// CHECK11-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 3
 // CHECK11-NEXT:    [[TMP242:%.*]] = bitcast i8** [[TMP241]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP218]], i32** [[TMP242]], align 4
-// CHECK11-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 4
+// CHECK11-NEXT:    store i32* [[TMP223]], i32** [[TMP242]], align 4
+// CHECK11-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 3
 // CHECK11-NEXT:    store i8* null, i8** [[TMP243]], align 4
-// CHECK11-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP246:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_75]], align 4
-// CHECK11-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_75]], align 4
-// CHECK11-NEXT:    [[SUB77:%.*]] = sub nsw i32 [[TMP247]], 0
-// CHECK11-NEXT:    [[DIV78:%.*]] = sdiv i32 [[SUB77]], 1
-// CHECK11-NEXT:    [[SUB79:%.*]] = sub nsw i32 [[DIV78]], 1
-// CHECK11-NEXT:    store i32 [[SUB79]], i32* [[DOTCAPTURE_EXPR_76]], align 4
-// CHECK11-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_76]], align 4
-// CHECK11-NEXT:    [[ADD80:%.*]] = add nsw i32 [[TMP248]], 1
-// CHECK11-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD80]] to i64
+// CHECK11-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_65]], align 4
+// CHECK11-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_65]], align 4
+// CHECK11-NEXT:    [[SUB67:%.*]] = sub nsw i32 [[TMP247]], 0
+// CHECK11-NEXT:    [[DIV68:%.*]] = sdiv i32 [[SUB67]], 1
+// CHECK11-NEXT:    [[SUB69:%.*]] = sub nsw i32 [[DIV68]], 1
+// CHECK11-NEXT:    store i32 [[SUB69]], i32* [[DOTCAPTURE_EXPR_66]], align 4
+// CHECK11-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_66]], align 4
+// CHECK11-NEXT:    [[ADD70:%.*]] = add nsw i32 [[TMP248]], 1
+// CHECK11-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD70]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP249]])
-// CHECK11-NEXT:    [[TMP250:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l91.region_id, i32 5, i8** [[TMP244]], i8** [[TMP245]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.52, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.53, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP251:%.*]] = icmp ne i32 [[TMP250]], 0
-// CHECK11-NEXT:    br i1 [[TMP251]], label [[OMP_OFFLOAD_FAILED81:%.*]], label [[OMP_OFFLOAD_CONT82:%.*]]
-// CHECK11:       omp_offload.failed81:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l91(i32 [[TMP213]], i32 [[TMP215]], i32* [[TMP216]], i32* [[TMP217]], i32* [[TMP218]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT82]]
-// CHECK11:       omp_offload.cont82:
+// CHECK11-NEXT:    [[KERNEL_ARGS71:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP250:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP250]], align 4
+// CHECK11-NEXT:    [[TMP251:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP251]], align 4
+// CHECK11-NEXT:    [[TMP252:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP244]], i8*** [[TMP252]], align 4
+// CHECK11-NEXT:    [[TMP253:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP245]], i8*** [[TMP253]], align 4
+// CHECK11-NEXT:    [[TMP254:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.48, i32 0, i32 0), i64** [[TMP254]], align 4
+// CHECK11-NEXT:    [[TMP255:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.49, i32 0, i32 0), i64** [[TMP255]], align 4
+// CHECK11-NEXT:    [[TMP256:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP256]], align 4
+// CHECK11-NEXT:    [[TMP257:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP257]], align 4
+// CHECK11-NEXT:    [[TMP258:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l83.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]])
+// CHECK11-NEXT:    [[TMP259:%.*]] = icmp ne i32 [[TMP258]], 0
+// CHECK11-NEXT:    br i1 [[TMP259]], label [[OMP_OFFLOAD_FAILED72:%.*]], label [[OMP_OFFLOAD_CONT73:%.*]]
+// CHECK11:       omp_offload.failed72:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l83(i32 [[TMP220]], i32* [[TMP221]], i32* [[TMP222]], i32* [[TMP223]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT73]]
+// CHECK11:       omp_offload.cont73:
+// CHECK11-NEXT:    [[TMP260:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK11-NEXT:    store i32 [[TMP260]], i32* [[CH_CASTED74]], align 4
+// CHECK11-NEXT:    [[TMP261:%.*]] = load i32, i32* [[CH_CASTED74]], align 4
+// CHECK11-NEXT:    [[TMP262:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP262]], i32* [[N_CASTED75]], align 4
+// CHECK11-NEXT:    [[TMP263:%.*]] = load i32, i32* [[N_CASTED75]], align 4
+// CHECK11-NEXT:    [[TMP264:%.*]] = load i32*, i32** [[A]], align 4
+// CHECK11-NEXT:    [[TMP265:%.*]] = load i32*, i32** [[B]], align 4
+// CHECK11-NEXT:    [[TMP266:%.*]] = load i32*, i32** [[C]], align 4
+// CHECK11-NEXT:    [[TMP267:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP268:%.*]] = bitcast i8** [[TMP267]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP261]], i32* [[TMP268]], align 4
+// CHECK11-NEXT:    [[TMP269:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP270:%.*]] = bitcast i8** [[TMP269]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP261]], i32* [[TMP270]], align 4
+// CHECK11-NEXT:    [[TMP271:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP271]], align 4
+// CHECK11-NEXT:    [[TMP272:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP273:%.*]] = bitcast i8** [[TMP272]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP263]], i32* [[TMP273]], align 4
+// CHECK11-NEXT:    [[TMP274:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP275:%.*]] = bitcast i8** [[TMP274]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP263]], i32* [[TMP275]], align 4
+// CHECK11-NEXT:    [[TMP276:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP276]], align 4
+// CHECK11-NEXT:    [[TMP277:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP278:%.*]] = bitcast i8** [[TMP277]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP264]], i32** [[TMP278]], align 4
+// CHECK11-NEXT:    [[TMP279:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP280:%.*]] = bitcast i8** [[TMP279]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP264]], i32** [[TMP280]], align 4
+// CHECK11-NEXT:    [[TMP281:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP281]], align 4
+// CHECK11-NEXT:    [[TMP282:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP283:%.*]] = bitcast i8** [[TMP282]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP265]], i32** [[TMP283]], align 4
+// CHECK11-NEXT:    [[TMP284:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP285:%.*]] = bitcast i8** [[TMP284]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP265]], i32** [[TMP285]], align 4
+// CHECK11-NEXT:    [[TMP286:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP286]], align 4
+// CHECK11-NEXT:    [[TMP287:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP288:%.*]] = bitcast i8** [[TMP287]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP266]], i32** [[TMP288]], align 4
+// CHECK11-NEXT:    [[TMP289:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP290:%.*]] = bitcast i8** [[TMP289]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP266]], i32** [[TMP290]], align 4
+// CHECK11-NEXT:    [[TMP291:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 4
+// CHECK11-NEXT:    store i8* null, i8** [[TMP291]], align 4
+// CHECK11-NEXT:    [[TMP292:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP293:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP294:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP294]], i32* [[DOTCAPTURE_EXPR_80]], align 4
+// CHECK11-NEXT:    [[TMP295:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_80]], align 4
+// CHECK11-NEXT:    [[SUB82:%.*]] = sub nsw i32 [[TMP295]], 0
+// CHECK11-NEXT:    [[DIV83:%.*]] = sdiv i32 [[SUB82]], 1
+// CHECK11-NEXT:    [[SUB84:%.*]] = sub nsw i32 [[DIV83]], 1
+// CHECK11-NEXT:    store i32 [[SUB84]], i32* [[DOTCAPTURE_EXPR_81]], align 4
+// CHECK11-NEXT:    [[TMP296:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_81]], align 4
+// CHECK11-NEXT:    [[ADD85:%.*]] = add nsw i32 [[TMP296]], 1
+// CHECK11-NEXT:    [[TMP297:%.*]] = zext i32 [[ADD85]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP297]])
+// CHECK11-NEXT:    [[KERNEL_ARGS86:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP298:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP298]], align 4
+// CHECK11-NEXT:    [[TMP299:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP299]], align 4
+// CHECK11-NEXT:    [[TMP300:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP292]], i8*** [[TMP300]], align 4
+// CHECK11-NEXT:    [[TMP301:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP293]], i8*** [[TMP301]], align 4
+// CHECK11-NEXT:    [[TMP302:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.52, i32 0, i32 0), i64** [[TMP302]], align 4
+// CHECK11-NEXT:    [[TMP303:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.53, i32 0, i32 0), i64** [[TMP303]], align 4
+// CHECK11-NEXT:    [[TMP304:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP304]], align 4
+// CHECK11-NEXT:    [[TMP305:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP305]], align 4
+// CHECK11-NEXT:    [[TMP306:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l91.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]])
+// CHECK11-NEXT:    [[TMP307:%.*]] = icmp ne i32 [[TMP306]], 0
+// CHECK11-NEXT:    br i1 [[TMP307]], label [[OMP_OFFLOAD_FAILED87:%.*]], label [[OMP_OFFLOAD_CONT88:%.*]]
+// CHECK11:       omp_offload.failed87:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l91(i32 [[TMP261]], i32 [[TMP263]], i32* [[TMP264]], i32* [[TMP265]], i32* [[TMP266]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT88]]
+// CHECK11:       omp_offload.cont88:
 // CHECK11-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/distribute_parallel_for_firstprivate_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_firstprivate_codegen.cpp
index bf9ab60..508dbb6 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_firstprivate_codegen.cpp
@@ -868,9 +868,26 @@ int main() {
 // CHECK8-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK8-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK8-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK8-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK8-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK8-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK8-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK8-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK8-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK8-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK8-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK8-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK8-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 8
+// CHECK8-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK8-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 8
+// CHECK8-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK8-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK8-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK8-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK8-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK8-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK8-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK8-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK8-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK8-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK8-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK8:       omp_offload.failed:
 // CHECK8-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK8-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -878,18 +895,18 @@ int main() {
 // CHECK8-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK8-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK8-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK8-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK8-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK8-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK8:       arraydestroy.body:
-// CHECK8-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK8-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK8-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK8-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK8-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK8-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK8:       arraydestroy.done3:
 // CHECK8-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK8-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK8-NEXT:    ret i32 [[TMP39]]
+// CHECK8-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK8-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK8-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1300,27 +1317,44 @@ int main() {
 // CHECK8-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK8-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK8-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK8-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l48.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK8-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK8-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK8-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK8-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK8-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK8-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK8-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK8-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK8-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK8-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK8-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK8-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK8-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK8-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK8-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK8-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK8-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK8-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK8-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK8-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK8-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK8-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK8:       omp_offload.failed:
 // CHECK8-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l48(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK8-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK8:       omp_offload.cont:
 // CHECK8-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK8-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK8-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK8-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK8-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK8:       arraydestroy.body:
-// CHECK8-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK8-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK8-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK8-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK8-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK8-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK8:       arraydestroy.done2:
 // CHECK8-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK8-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK8-NEXT:    ret i32 [[TMP32]]
+// CHECK8-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK8-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK8-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1804,9 +1838,26 @@ int main() {
 // CHECK10-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK10-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK10-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK10-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK10-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK10-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK10-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK10-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK10-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK10-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK10-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK10-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK10-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 4
+// CHECK10-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK10-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 4
+// CHECK10-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK10-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK10-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK10-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK10-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK10-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK10-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK10-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK10-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK10-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK10-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK10:       omp_offload.failed:
 // CHECK10-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK10-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1814,18 +1865,18 @@ int main() {
 // CHECK10-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK10-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK10-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK10-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK10-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK10-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK10:       arraydestroy.body:
-// CHECK10-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK10-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK10-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK10-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK10-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK10-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK10:       arraydestroy.done2:
 // CHECK10-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK10-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK10-NEXT:    ret i32 [[TMP39]]
+// CHECK10-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK10-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK10-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -2223,27 +2274,44 @@ int main() {
 // CHECK10-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK10-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK10-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK10-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l48.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK10-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK10-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK10-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK10-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK10-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK10-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK10-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK10-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK10-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK10-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK10-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK10-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK10-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK10-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK10-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK10-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK10-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK10-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK10-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK10-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK10-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK10-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK10:       omp_offload.failed:
 // CHECK10-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l48(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK10-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK10:       omp_offload.cont:
 // CHECK10-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK10-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK10-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK10-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK10-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK10:       arraydestroy.body:
-// CHECK10-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK10-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK10-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK10-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK10-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK10-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK10:       arraydestroy.done2:
 // CHECK10-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK10-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK10-NEXT:    ret i32 [[TMP32]]
+// CHECK10-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK10-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK10-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/distribute_parallel_for_if_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_if_codegen.cpp
index d67036e..3718fc9 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_if_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_if_codegen.cpp
@@ -126,21 +126,55 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47() #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret void
 //
 //
@@ -434,48 +468,99 @@ int main() {
 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
+// CHECK1-NEXT:    [[TMP20:%.*]] = load i32, i32* @Arg, align 4
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK1-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK1:       omp_offload.failed5:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK1:       omp_offload.cont6:
-// CHECK1-NEXT:    [[TMP15:%.*]] = load i32, i32* @Arg, align 4
-// CHECK1-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP15]])
+// CHECK1-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK1-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK1-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l103(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK1:       omp_offload.cont8:
+// CHECK1-NEXT:    [[TMP39:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP39]])
 // CHECK1-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -928,46 +1013,97 @@ int main() {
 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l63() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l69() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
+// CHECK1-NEXT:    [[TMP20:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK1-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK1:       omp_offload.failed5:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK1:       omp_offload.cont6:
+// CHECK1-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK1-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK1-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l75(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK1:       omp_offload.cont8:
 // CHECK1-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/distribute_parallel_for_lastprivate_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_lastprivate_codegen.cpp
index 2664830..e8c88d0 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_lastprivate_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_lastprivate_codegen.cpp
@@ -832,9 +832,26 @@ int main() {
 // CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l123.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK9-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l123.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK9-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l123(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -842,18 +859,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done3:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP39]]
+// CHECK9-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1296,27 +1313,44 @@ int main() {
 // CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK9-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK9-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP32]]
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1835,9 +1869,26 @@ int main() {
 // CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l123.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK11-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l123.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK11-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l123(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1845,18 +1896,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP39]]
+// CHECK11-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -2290,27 +2341,44 @@ int main() {
 // CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK11-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK11-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP32]]
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/distribute_parallel_for_num_threads_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_num_threads_codegen.cpp
index 346c8cf..c4fc339 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_num_threads_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_num_threads_codegen.cpp
@@ -114,64 +114,98 @@ int main() {
 // CHECK1:       invoke.cont:
 // CHECK1-NEXT:    store i8 [[CALL]], i8* [[A]], align 1
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68() #[[ATTR6:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       lpad:
-// CHECK1-NEXT:    [[TMP2:%.*]] = landingpad { i8*, i32 }
+// CHECK1-NEXT:    [[TMP10:%.*]] = landingpad { i8*, i32 }
 // CHECK1-NEXT:    cleanup
-// CHECK1-NEXT:    [[TMP3:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 0
-// CHECK1-NEXT:    store i8* [[TMP3]], i8** [[EXN_SLOT]], align 8
-// CHECK1-NEXT:    [[TMP4:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 1
-// CHECK1-NEXT:    store i32 [[TMP4]], i32* [[EHSELECTOR_SLOT]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 0
+// CHECK1-NEXT:    store i8* [[TMP11]], i8** [[EXN_SLOT]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 1
+// CHECK1-NEXT:    store i32 [[TMP12]], i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK1-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
 // CHECK1-NEXT:    br label [[EH_RESUME:%.*]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i8, i8* [[A]], align 1
+// CHECK1-NEXT:    [[TMP13:%.*]] = load i8, i8* [[A]], align 1
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i8*
-// CHECK1-NEXT:    store i8 [[TMP5]], i8* [[CONV]], align 1
-// CHECK1-NEXT:    [[TMP6:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP8]], align 8
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP11]], align 8
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    store i8 [[TMP13]], i8* [[CONV]], align 1
+// CHECK1-NEXT:    [[TMP14:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, i32 1, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK1-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP6]]) #[[ATTR6]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
-// CHECK1-NEXT:    [[TMP16:%.*]] = load i8, i8* [[A]], align 1
-// CHECK1-NEXT:    [[CONV4:%.*]] = sext i8 [[TMP16]] to i32
-// CHECK1-NEXT:    [[CALL6:%.*]] = invoke noundef i32 @_Z5tmainIcLi5EEiv()
-// CHECK1-NEXT:    to label [[INVOKE_CONT5:%.*]] unwind label [[LPAD]]
-// CHECK1:       invoke.cont5:
-// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV4]], [[CALL6]]
-// CHECK1-NEXT:    [[CALL8:%.*]] = invoke noundef i32 @_Z5tmainI1SLi1EEiv()
-// CHECK1-NEXT:    to label [[INVOKE_CONT7:%.*]] unwind label [[LPAD]]
-// CHECK1:       invoke.cont7:
-// CHECK1-NEXT:    [[ADD9:%.*]] = add nsw i32 [[ADD]], [[CALL8]]
-// CHECK1-NEXT:    store i32 [[ADD9]], i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK1-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP14]]) #[[ATTR6]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
+// CHECK1-NEXT:    [[TMP32:%.*]] = load i8, i8* [[A]], align 1
+// CHECK1-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP32]] to i32
+// CHECK1-NEXT:    [[CALL7:%.*]] = invoke noundef i32 @_Z5tmainIcLi5EEiv()
+// CHECK1-NEXT:    to label [[INVOKE_CONT6:%.*]] unwind label [[LPAD]]
+// CHECK1:       invoke.cont6:
+// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV5]], [[CALL7]]
+// CHECK1-NEXT:    [[CALL9:%.*]] = invoke noundef i32 @_Z5tmainI1SLi1EEiv()
+// CHECK1-NEXT:    to label [[INVOKE_CONT8:%.*]] unwind label [[LPAD]]
+// CHECK1:       invoke.cont8:
+// CHECK1-NEXT:    [[ADD10:%.*]] = add nsw i32 [[ADD]], [[CALL9]]
+// CHECK1-NEXT:    store i32 [[ADD10]], i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
-// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP17]]
+// CHECK1-NEXT:    [[TMP33:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP33]]
 // CHECK1:       eh.resume:
 // CHECK1-NEXT:    [[EXN:%.*]] = load i8*, i8** [[EXN_SLOT]], align 8
 // CHECK1-NEXT:    [[SEL:%.*]] = load i32, i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK1-NEXT:    [[LPAD_VAL:%.*]] = insertvalue { i8*, i32 } undef, i8* [[EXN]], 0
-// CHECK1-NEXT:    [[LPAD_VAL10:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
-// CHECK1-NEXT:    resume { i8*, i32 } [[LPAD_VAL10]]
+// CHECK1-NEXT:    [[LPAD_VAL11:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
+// CHECK1-NEXT:    resume { i8*, i32 } [[LPAD_VAL11]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN1SC1El
@@ -515,21 +549,55 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52() #[[ATTR6]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57() #[[ATTR6]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret i32 0
 //
 //
@@ -539,21 +607,55 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52() #[[ATTR6]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57() #[[ATTR6]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret i32 0
 //
 //
@@ -1221,64 +1323,98 @@ int main() {
 // CHECK5:       invoke.cont:
 // CHECK5-NEXT:    store i8 [[CALL]], i8* [[A]], align 1
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK5-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK5-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK5-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK5-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK5-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK5-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68() #[[ATTR6:[0-9]+]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       lpad:
-// CHECK5-NEXT:    [[TMP2:%.*]] = landingpad { i8*, i32 }
+// CHECK5-NEXT:    [[TMP10:%.*]] = landingpad { i8*, i32 }
 // CHECK5-NEXT:    cleanup
-// CHECK5-NEXT:    [[TMP3:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 0
-// CHECK5-NEXT:    store i8* [[TMP3]], i8** [[EXN_SLOT]], align 8
-// CHECK5-NEXT:    [[TMP4:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 1
-// CHECK5-NEXT:    store i32 [[TMP4]], i32* [[EHSELECTOR_SLOT]], align 4
+// CHECK5-NEXT:    [[TMP11:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 0
+// CHECK5-NEXT:    store i8* [[TMP11]], i8** [[EXN_SLOT]], align 8
+// CHECK5-NEXT:    [[TMP12:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 1
+// CHECK5-NEXT:    store i32 [[TMP12]], i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK5-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
 // CHECK5-NEXT:    br label [[EH_RESUME:%.*]]
 // CHECK5:       omp_offload.cont:
-// CHECK5-NEXT:    [[TMP5:%.*]] = load i8, i8* [[A]], align 1
+// CHECK5-NEXT:    [[TMP13:%.*]] = load i8, i8* [[A]], align 1
 // CHECK5-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i8*
-// CHECK5-NEXT:    store i8 [[TMP5]], i8* [[CONV]], align 1
-// CHECK5-NEXT:    [[TMP6:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP6]], i64* [[TMP8]], align 8
-// CHECK5-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP6]], i64* [[TMP10]], align 8
-// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP11]], align 8
-// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK5-NEXT:    store i8 [[TMP13]], i8* [[CONV]], align 1
+// CHECK5-NEXT:    [[TMP14:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP14]], i64* [[TMP16]], align 8
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP14]], i64* [[TMP18]], align 8
+// CHECK5-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP19]], align 8
+// CHECK5-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, i32 1, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK5-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK5:       omp_offload.failed2:
-// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP6]]) #[[ATTR6]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK5:       omp_offload.cont3:
-// CHECK5-NEXT:    [[TMP16:%.*]] = load i8, i8* [[A]], align 1
-// CHECK5-NEXT:    [[CONV4:%.*]] = sext i8 [[TMP16]] to i32
-// CHECK5-NEXT:    [[CALL6:%.*]] = invoke noundef i32 @_Z5tmainIcLi5EEiv()
-// CHECK5-NEXT:    to label [[INVOKE_CONT5:%.*]] unwind label [[LPAD]]
-// CHECK5:       invoke.cont5:
-// CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV4]], [[CALL6]]
-// CHECK5-NEXT:    [[CALL8:%.*]] = invoke noundef i32 @_Z5tmainI1SLi1EEiv()
-// CHECK5-NEXT:    to label [[INVOKE_CONT7:%.*]] unwind label [[LPAD]]
-// CHECK5:       invoke.cont7:
-// CHECK5-NEXT:    [[ADD9:%.*]] = add nsw i32 [[ADD]], [[CALL8]]
-// CHECK5-NEXT:    store i32 [[ADD9]], i32* [[RETVAL]], align 4
+// CHECK5-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK5-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK5-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK5-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK5:       omp_offload.failed3:
+// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP14]]) #[[ATTR6]]
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK5:       omp_offload.cont4:
+// CHECK5-NEXT:    [[TMP32:%.*]] = load i8, i8* [[A]], align 1
+// CHECK5-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP32]] to i32
+// CHECK5-NEXT:    [[CALL7:%.*]] = invoke noundef i32 @_Z5tmainIcLi5EEiv()
+// CHECK5-NEXT:    to label [[INVOKE_CONT6:%.*]] unwind label [[LPAD]]
+// CHECK5:       invoke.cont6:
+// CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV5]], [[CALL7]]
+// CHECK5-NEXT:    [[CALL9:%.*]] = invoke noundef i32 @_Z5tmainI1SLi1EEiv()
+// CHECK5-NEXT:    to label [[INVOKE_CONT8:%.*]] unwind label [[LPAD]]
+// CHECK5:       invoke.cont8:
+// CHECK5-NEXT:    [[ADD10:%.*]] = add nsw i32 [[ADD]], [[CALL9]]
+// CHECK5-NEXT:    store i32 [[ADD10]], i32* [[RETVAL]], align 4
 // CHECK5-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
-// CHECK5-NEXT:    [[TMP17:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP17]]
+// CHECK5-NEXT:    [[TMP33:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP33]]
 // CHECK5:       eh.resume:
 // CHECK5-NEXT:    [[EXN:%.*]] = load i8*, i8** [[EXN_SLOT]], align 8
 // CHECK5-NEXT:    [[SEL:%.*]] = load i32, i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK5-NEXT:    [[LPAD_VAL:%.*]] = insertvalue { i8*, i32 } undef, i8* [[EXN]], 0
-// CHECK5-NEXT:    [[LPAD_VAL10:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
-// CHECK5-NEXT:    resume { i8*, i32 } [[LPAD_VAL10]]
+// CHECK5-NEXT:    [[LPAD_VAL11:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
+// CHECK5-NEXT:    resume { i8*, i32 } [[LPAD_VAL11]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@_ZN1SC1El
@@ -1622,21 +1758,55 @@ int main() {
 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK5-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK5-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK5-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK5-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK5-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK5-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52() #[[ATTR6]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK5-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK5:       omp_offload.failed2:
+// CHECK5-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK5-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK5-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK5-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK5-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK5:       omp_offload.failed3:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57() #[[ATTR6]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK5:       omp_offload.cont3:
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK5:       omp_offload.cont4:
 // CHECK5-NEXT:    ret i32 0
 //
 //
@@ -1646,21 +1816,55 @@ int main() {
 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK5-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK5-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK5-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK5-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK5-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK5-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52() #[[ATTR6]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK5-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK5:       omp_offload.failed2:
+// CHECK5-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK5-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK5-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK5-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK5-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK5:       omp_offload.failed3:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57() #[[ATTR6]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK5:       omp_offload.cont3:
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK5:       omp_offload.cont4:
 // CHECK5-NEXT:    ret i32 0
 //
 //
@@ -2328,64 +2532,98 @@ int main() {
 // CHECK9:       invoke.cont:
 // CHECK9-NEXT:    store i8 [[CALL]], i8* [[A]], align 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68() #[[ATTR6:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       lpad:
-// CHECK9-NEXT:    [[TMP2:%.*]] = landingpad { i8*, i32 }
+// CHECK9-NEXT:    [[TMP10:%.*]] = landingpad { i8*, i32 }
 // CHECK9-NEXT:    cleanup
-// CHECK9-NEXT:    [[TMP3:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 0
-// CHECK9-NEXT:    store i8* [[TMP3]], i8** [[EXN_SLOT]], align 8
-// CHECK9-NEXT:    [[TMP4:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 1
-// CHECK9-NEXT:    store i32 [[TMP4]], i32* [[EHSELECTOR_SLOT]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 0
+// CHECK9-NEXT:    store i8* [[TMP11]], i8** [[EXN_SLOT]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 1
+// CHECK9-NEXT:    store i32 [[TMP12]], i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK9-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
 // CHECK9-NEXT:    br label [[EH_RESUME:%.*]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP5:%.*]] = load i8, i8* [[A]], align 1
+// CHECK9-NEXT:    [[TMP13:%.*]] = load i8, i8* [[A]], align 1
 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i8*
-// CHECK9-NEXT:    store i8 [[TMP5]], i8* [[CONV]], align 1
-// CHECK9-NEXT:    [[TMP6:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP6]], i64* [[TMP8]], align 8
-// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP6]], i64* [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP11]], align 8
-// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    store i8 [[TMP13]], i8* [[CONV]], align 1
+// CHECK9-NEXT:    [[TMP14:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP14]], i64* [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP14]], i64* [[TMP18]], align 8
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP19]], align 8
+// CHECK9-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, i32 1, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK9-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK9:       omp_offload.failed2:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP6]]) #[[ATTR6]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK9:       omp_offload.cont3:
-// CHECK9-NEXT:    [[TMP16:%.*]] = load i8, i8* [[A]], align 1
-// CHECK9-NEXT:    [[CONV4:%.*]] = sext i8 [[TMP16]] to i32
-// CHECK9-NEXT:    [[CALL6:%.*]] = invoke noundef i32 @_Z5tmainIcLi5EEiv()
-// CHECK9-NEXT:    to label [[INVOKE_CONT5:%.*]] unwind label [[LPAD]]
-// CHECK9:       invoke.cont5:
-// CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV4]], [[CALL6]]
-// CHECK9-NEXT:    [[CALL8:%.*]] = invoke noundef i32 @_Z5tmainI1SLi1EEiv()
-// CHECK9-NEXT:    to label [[INVOKE_CONT7:%.*]] unwind label [[LPAD]]
-// CHECK9:       invoke.cont7:
-// CHECK9-NEXT:    [[ADD9:%.*]] = add nsw i32 [[ADD]], [[CALL8]]
-// CHECK9-NEXT:    store i32 [[ADD9]], i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK9:       omp_offload.failed3:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP14]]) #[[ATTR6]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK9:       omp_offload.cont4:
+// CHECK9-NEXT:    [[TMP32:%.*]] = load i8, i8* [[A]], align 1
+// CHECK9-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP32]] to i32
+// CHECK9-NEXT:    [[CALL7:%.*]] = invoke noundef i32 @_Z5tmainIcLi5EEiv()
+// CHECK9-NEXT:    to label [[INVOKE_CONT6:%.*]] unwind label [[LPAD]]
+// CHECK9:       invoke.cont6:
+// CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV5]], [[CALL7]]
+// CHECK9-NEXT:    [[CALL9:%.*]] = invoke noundef i32 @_Z5tmainI1SLi1EEiv()
+// CHECK9-NEXT:    to label [[INVOKE_CONT8:%.*]] unwind label [[LPAD]]
+// CHECK9:       invoke.cont8:
+// CHECK9-NEXT:    [[ADD10:%.*]] = add nsw i32 [[ADD]], [[CALL9]]
+// CHECK9-NEXT:    store i32 [[ADD10]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
-// CHECK9-NEXT:    [[TMP17:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP17]]
+// CHECK9-NEXT:    [[TMP33:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP33]]
 // CHECK9:       eh.resume:
 // CHECK9-NEXT:    [[EXN:%.*]] = load i8*, i8** [[EXN_SLOT]], align 8
 // CHECK9-NEXT:    [[SEL:%.*]] = load i32, i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK9-NEXT:    [[LPAD_VAL:%.*]] = insertvalue { i8*, i32 } undef, i8* [[EXN]], 0
-// CHECK9-NEXT:    [[LPAD_VAL10:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
-// CHECK9-NEXT:    resume { i8*, i32 } [[LPAD_VAL10]]
+// CHECK9-NEXT:    [[LPAD_VAL11:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
+// CHECK9-NEXT:    resume { i8*, i32 } [[LPAD_VAL11]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SC1El
@@ -2729,21 +2967,55 @@ int main() {
 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52() #[[ATTR6]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK9-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK9:       omp_offload.failed2:
+// CHECK9-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK9-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK9-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK9:       omp_offload.failed3:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57() #[[ATTR6]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK9:       omp_offload.cont3:
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK9:       omp_offload.cont4:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -2753,21 +3025,55 @@ int main() {
 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52() #[[ATTR6]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK9-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK9:       omp_offload.failed2:
+// CHECK9-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK9-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK9-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK9:       omp_offload.failed3:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57() #[[ATTR6]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK9:       omp_offload.cont3:
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK9:       omp_offload.cont4:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -3435,64 +3741,98 @@ int main() {
 // CHECK13:       invoke.cont:
 // CHECK13-NEXT:    store i8 [[CALL]], i8* [[A]], align 1
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK13-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK13-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK13-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK13-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK13-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK13-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK13-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK13-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK13-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK13-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK13-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68() #[[ATTR6:[0-9]+]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       lpad:
-// CHECK13-NEXT:    [[TMP2:%.*]] = landingpad { i8*, i32 }
+// CHECK13-NEXT:    [[TMP10:%.*]] = landingpad { i8*, i32 }
 // CHECK13-NEXT:    cleanup
-// CHECK13-NEXT:    [[TMP3:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 0
-// CHECK13-NEXT:    store i8* [[TMP3]], i8** [[EXN_SLOT]], align 8
-// CHECK13-NEXT:    [[TMP4:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 1
-// CHECK13-NEXT:    store i32 [[TMP4]], i32* [[EHSELECTOR_SLOT]], align 4
+// CHECK13-NEXT:    [[TMP11:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 0
+// CHECK13-NEXT:    store i8* [[TMP11]], i8** [[EXN_SLOT]], align 8
+// CHECK13-NEXT:    [[TMP12:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 1
+// CHECK13-NEXT:    store i32 [[TMP12]], i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK13-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
 // CHECK13-NEXT:    br label [[EH_RESUME:%.*]]
 // CHECK13:       omp_offload.cont:
-// CHECK13-NEXT:    [[TMP5:%.*]] = load i8, i8* [[A]], align 1
+// CHECK13-NEXT:    [[TMP13:%.*]] = load i8, i8* [[A]], align 1
 // CHECK13-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i8*
-// CHECK13-NEXT:    store i8 [[TMP5]], i8* [[CONV]], align 1
-// CHECK13-NEXT:    [[TMP6:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK13-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP6]], i64* [[TMP8]], align 8
-// CHECK13-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP6]], i64* [[TMP10]], align 8
-// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP11]], align 8
-// CHECK13-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK13-NEXT:    store i8 [[TMP13]], i8* [[CONV]], align 1
+// CHECK13-NEXT:    [[TMP14:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK13-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP14]], i64* [[TMP16]], align 8
+// CHECK13-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP14]], i64* [[TMP18]], align 8
+// CHECK13-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP19]], align 8
+// CHECK13-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK13-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, i32 1, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK13-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK13:       omp_offload.failed2:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP6]]) #[[ATTR6]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK13:       omp_offload.cont3:
-// CHECK13-NEXT:    [[TMP16:%.*]] = load i8, i8* [[A]], align 1
-// CHECK13-NEXT:    [[CONV4:%.*]] = sext i8 [[TMP16]] to i32
-// CHECK13-NEXT:    [[CALL6:%.*]] = invoke noundef i32 @_Z5tmainIcLi5EEiv()
-// CHECK13-NEXT:    to label [[INVOKE_CONT5:%.*]] unwind label [[LPAD]]
-// CHECK13:       invoke.cont5:
-// CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV4]], [[CALL6]]
-// CHECK13-NEXT:    [[CALL8:%.*]] = invoke noundef i32 @_Z5tmainI1SLi1EEiv()
-// CHECK13-NEXT:    to label [[INVOKE_CONT7:%.*]] unwind label [[LPAD]]
-// CHECK13:       invoke.cont7:
-// CHECK13-NEXT:    [[ADD9:%.*]] = add nsw i32 [[ADD]], [[CALL8]]
-// CHECK13-NEXT:    store i32 [[ADD9]], i32* [[RETVAL]], align 4
+// CHECK13-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK13-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK13-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK13-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK13-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK13-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK13-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK13-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK13-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK13-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK13-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK13:       omp_offload.failed3:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP14]]) #[[ATTR6]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK13:       omp_offload.cont4:
+// CHECK13-NEXT:    [[TMP32:%.*]] = load i8, i8* [[A]], align 1
+// CHECK13-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP32]] to i32
+// CHECK13-NEXT:    [[CALL7:%.*]] = invoke noundef i32 @_Z5tmainIcLi5EEiv()
+// CHECK13-NEXT:    to label [[INVOKE_CONT6:%.*]] unwind label [[LPAD]]
+// CHECK13:       invoke.cont6:
+// CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV5]], [[CALL7]]
+// CHECK13-NEXT:    [[CALL9:%.*]] = invoke noundef i32 @_Z5tmainI1SLi1EEiv()
+// CHECK13-NEXT:    to label [[INVOKE_CONT8:%.*]] unwind label [[LPAD]]
+// CHECK13:       invoke.cont8:
+// CHECK13-NEXT:    [[ADD10:%.*]] = add nsw i32 [[ADD]], [[CALL9]]
+// CHECK13-NEXT:    store i32 [[ADD10]], i32* [[RETVAL]], align 4
 // CHECK13-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
-// CHECK13-NEXT:    [[TMP17:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK13-NEXT:    ret i32 [[TMP17]]
+// CHECK13-NEXT:    [[TMP33:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK13-NEXT:    ret i32 [[TMP33]]
 // CHECK13:       eh.resume:
 // CHECK13-NEXT:    [[EXN:%.*]] = load i8*, i8** [[EXN_SLOT]], align 8
 // CHECK13-NEXT:    [[SEL:%.*]] = load i32, i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK13-NEXT:    [[LPAD_VAL:%.*]] = insertvalue { i8*, i32 } undef, i8* [[EXN]], 0
-// CHECK13-NEXT:    [[LPAD_VAL10:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
-// CHECK13-NEXT:    resume { i8*, i32 } [[LPAD_VAL10]]
+// CHECK13-NEXT:    [[LPAD_VAL11:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
+// CHECK13-NEXT:    resume { i8*, i32 } [[LPAD_VAL11]]
 //
 //
 // CHECK13-LABEL: define {{[^@]+}}@_ZN1SC1El
@@ -3836,21 +4176,55 @@ int main() {
 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK13-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK13-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK13-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK13-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK13-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK13-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK13-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK13-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK13-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK13-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK13-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52() #[[ATTR6]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       omp_offload.cont:
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK13-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK13-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK13:       omp_offload.failed2:
+// CHECK13-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK13-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK13-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK13-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK13-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK13-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK13-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK13-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK13-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK13:       omp_offload.failed3:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57() #[[ATTR6]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK13:       omp_offload.cont3:
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK13:       omp_offload.cont4:
 // CHECK13-NEXT:    ret i32 0
 //
 //
@@ -3860,21 +4234,55 @@ int main() {
 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK13-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK13-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK13-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK13-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK13-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK13-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK13-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK13-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK13-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK13-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK13-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52() #[[ATTR6]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       omp_offload.cont:
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK13-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK13-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK13:       omp_offload.failed2:
+// CHECK13-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK13-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK13-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK13-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK13-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK13-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK13-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK13-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK13-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK13:       omp_offload.failed3:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57() #[[ATTR6]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK13:       omp_offload.cont3:
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK13:       omp_offload.cont4:
 // CHECK13-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/distribute_parallel_for_private_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_private_codegen.cpp
index 528b70c..25e889e 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_private_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_private_codegen.cpp
@@ -535,9 +535,26 @@ int main() {
 // CHECK9-NEXT:    store %struct.S* [[TEST]], %struct.S** [[VAR]], align 8
 // CHECK9-NEXT:    store %struct.S* undef, %struct.S** [[_TMP1]], align 8
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l95.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK9-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK9-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l95.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK9-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l95() #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -545,18 +562,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP4]]
+// CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -833,27 +850,44 @@ int main() {
 // CHECK9-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 8
 // CHECK9-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 8
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK9-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK9-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK9-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49() #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP4]]
+// CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1207,9 +1241,26 @@ int main() {
 // CHECK11-NEXT:    store %struct.S* [[TEST]], %struct.S** [[VAR]], align 4
 // CHECK11-NEXT:    store %struct.S* undef, %struct.S** [[_TMP1]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l95.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK11-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l95.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK11-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l95() #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1217,18 +1268,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP4]]
+// CHECK11-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1499,27 +1550,44 @@ int main() {
 // CHECK11-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 4
 // CHECK11-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK11-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK11-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49() #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP4]]
+// CHECK11-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/distribute_parallel_for_proc_bind_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_proc_bind_codegen.cpp
index 746ecc3..4d6570d 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_proc_bind_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_proc_bind_codegen.cpp
@@ -61,21 +61,55 @@ int main() {
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l37.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l37.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l37() #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l41.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l41.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l41() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK1-NEXT:    ret i32 [[CALL]]
 //
@@ -361,9 +395,26 @@ int main() {
 // CHECK1-NEXT:  entry:
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l29.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l29.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l29() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/distribute_parallel_for_reduction_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_reduction_codegen.cpp
index 34d328c..15aff90 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_reduction_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_reduction_codegen.cpp
@@ -63,7 +63,7 @@ int main() {
 }
 
 // CHECK-LABEL: main
-// CHECK: call{{.+}} @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}},
+// CHECK: call{{.+}} @__tgt_target_kernel(%struct.ident_t* @{{.+}},
 // CHECK: call void [[OFFL:@.+]](
 // CHECK: call{{.+}} [[TMAIN:@.+]](i{{32|64}}
 // CHECK: ret
@@ -82,7 +82,7 @@ int main() {
 // CHECK: ret void
 
 // CHECK: define{{.+}} [[TMAIN]](i{{32|64}}
-// CHECK: call{{.+}} @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}},
+// CHECK: call{{.+}} @__tgt_target_kernel(%struct.ident_t* @{{.+}},
 // CHECK: call void [[TOFFL:@.+]](
 // CHECK: ret
 

diff --git a/clang/test/OpenMP/distribute_parallel_for_simd_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_simd_codegen.cpp
index 8736aa7..f41e467 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_simd_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_simd_codegen.cpp
@@ -4700,43 +4700,43 @@ int main() {
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[CH_CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED18:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED32:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS34:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS35:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS36:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP37:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_38:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_39:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[CH_CASTED46:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED48:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS50:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS51:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS52:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP53:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_54:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_55:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED62:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS64:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS65:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS66:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP67:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_68:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_69:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[CH_CASTED76:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED78:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS80:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS81:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS82:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP83:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_84:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_85:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED34:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS36:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS37:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS38:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP39:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_40:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_41:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[CH_CASTED49:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[N_CASTED51:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS53:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS54:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS55:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP56:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_57:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_58:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED66:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS68:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS69:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS70:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP71:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_72:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_73:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[CH_CASTED81:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[N_CASTED83:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS85:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS86:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS87:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP88:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_89:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_90:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    store i32 10000, i32* [[N]], align 4
 // CHECK9-NEXT:    store i32 100, i32* [[CH]], align 4
@@ -4792,403 +4792,522 @@ int main() {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP29]], 1
 // CHECK9-NEXT:    [[TMP30:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP30]])
-// CHECK9-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l368.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK9-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP32]], align 4
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP25]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP26]], i8*** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l368.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK9-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l368(i64 [[TMP1]], double* [[TMP2]], double* [[TMP3]], double* [[TMP4]]) #[[ATTR2:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N]], align 4
 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP33]], i32* [[CONV4]], align 4
-// CHECK9-NEXT:    [[TMP34:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK9-NEXT:    [[TMP35:%.*]] = load double*, double** [[A]], align 8
-// CHECK9-NEXT:    [[TMP36:%.*]] = load double*, double** [[B]], align 8
-// CHECK9-NEXT:    [[TMP37:%.*]] = load double*, double** [[C]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP41]], align 8
-// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP42]], align 8
-// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to double**
-// CHECK9-NEXT:    store double* [[TMP35]], double** [[TMP44]], align 8
-// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to double**
-// CHECK9-NEXT:    store double* [[TMP35]], double** [[TMP46]], align 8
-// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP47]], align 8
-// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to double**
-// CHECK9-NEXT:    store double* [[TMP36]], double** [[TMP49]], align 8
-// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to double**
-// CHECK9-NEXT:    store double* [[TMP36]], double** [[TMP51]], align 8
-// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP52]], align 8
-// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 3
+// CHECK9-NEXT:    store i32 [[TMP41]], i32* [[CONV4]], align 4
+// CHECK9-NEXT:    [[TMP42:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = load double*, double** [[A]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = load double*, double** [[B]], align 8
+// CHECK9-NEXT:    [[TMP45:%.*]] = load double*, double** [[C]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP49]], align 8
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to double**
+// CHECK9-NEXT:    store double* [[TMP43]], double** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
 // CHECK9-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to double**
-// CHECK9-NEXT:    store double* [[TMP37]], double** [[TMP54]], align 8
-// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to double**
-// CHECK9-NEXT:    store double* [[TMP37]], double** [[TMP56]], align 8
-// CHECK9-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP57]], align 8
-// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP60:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP60]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK9-NEXT:    [[TMP61:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK9-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP61]], 0
+// CHECK9-NEXT:    store double* [[TMP43]], double** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to double**
+// CHECK9-NEXT:    store double* [[TMP44]], double** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to double**
+// CHECK9-NEXT:    store double* [[TMP44]], double** [[TMP59]], align 8
+// CHECK9-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP60]], align 8
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to double**
+// CHECK9-NEXT:    store double* [[TMP45]], double** [[TMP62]], align 8
+// CHECK9-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP64:%.*]] = bitcast i8** [[TMP63]] to double**
+// CHECK9-NEXT:    store double* [[TMP45]], double** [[TMP64]], align 8
+// CHECK9-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP65]], align 8
+// CHECK9-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP68]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK9-NEXT:    [[TMP69:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK9-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP69]], 0
 // CHECK9-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK9-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK9-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[TMP62:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP62]], 1
-// CHECK9-NEXT:    [[TMP63:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP63]])
-// CHECK9-NEXT:    [[TMP64:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l407.region_id, i32 4, i8** [[TMP58]], i8** [[TMP59]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP65:%.*]] = icmp ne i32 [[TMP64]], 0
-// CHECK9-NEXT:    br i1 [[TMP65]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK9:       omp_offload.failed15:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l407(i64 [[TMP34]], double* [[TMP35]], double* [[TMP36]], double* [[TMP37]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK9:       omp_offload.cont16:
-// CHECK9-NEXT:    [[TMP66:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK9-NEXT:    [[CONV17:%.*]] = bitcast i64* [[CH_CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP66]], i32* [[CONV17]], align 4
-// CHECK9-NEXT:    [[TMP67:%.*]] = load i64, i64* [[CH_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV19:%.*]] = bitcast i64* [[N_CASTED18]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP68]], i32* [[CONV19]], align 4
-// CHECK9-NEXT:    [[TMP69:%.*]] = load i64, i64* [[N_CASTED18]], align 8
-// CHECK9-NEXT:    [[TMP70:%.*]] = load double*, double** [[A]], align 8
-// CHECK9-NEXT:    [[TMP71:%.*]] = load double*, double** [[B]], align 8
-// CHECK9-NEXT:    [[TMP72:%.*]] = load double*, double** [[C]], align 8
-// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP67]], i64* [[TMP74]], align 8
-// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP67]], i64* [[TMP76]], align 8
-// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP77]], align 8
-// CHECK9-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP69]], i64* [[TMP79]], align 8
-// CHECK9-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP69]], i64* [[TMP81]], align 8
-// CHECK9-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP82]], align 8
-// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to double**
-// CHECK9-NEXT:    store double* [[TMP70]], double** [[TMP84]], align 8
-// CHECK9-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to double**
-// CHECK9-NEXT:    store double* [[TMP70]], double** [[TMP86]], align 8
-// CHECK9-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP87]], align 8
-// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to double**
-// CHECK9-NEXT:    store double* [[TMP71]], double** [[TMP89]], align 8
-// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to double**
-// CHECK9-NEXT:    store double* [[TMP71]], double** [[TMP91]], align 8
-// CHECK9-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP92]], align 8
-// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to double**
-// CHECK9-NEXT:    store double* [[TMP72]], double** [[TMP94]], align 8
-// CHECK9-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to double**
-// CHECK9-NEXT:    store double* [[TMP72]], double** [[TMP96]], align 8
-// CHECK9-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 4
-// CHECK9-NEXT:    store i8* null, i8** [[TMP97]], align 8
-// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP100]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK9-NEXT:    [[TMP101:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK9-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP101]], 0
-// CHECK9-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
-// CHECK9-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
-// CHECK9-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK9-NEXT:    [[TMP102:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK9-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP102]], 1
-// CHECK9-NEXT:    [[TMP103:%.*]] = zext i32 [[ADD29]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP103]])
-// CHECK9-NEXT:    [[TMP104:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l446.region_id, i32 5, i8** [[TMP98]], i8** [[TMP99]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP105:%.*]] = icmp ne i32 [[TMP104]], 0
-// CHECK9-NEXT:    br i1 [[TMP105]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK9:       omp_offload.failed30:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l446(i64 [[TMP67]], i64 [[TMP69]], double* [[TMP70]], double* [[TMP71]], double* [[TMP72]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK9:       omp_offload.cont31:
-// CHECK9-NEXT:    [[TMP106:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV33:%.*]] = bitcast i64* [[N_CASTED32]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP106]], i32* [[CONV33]], align 4
-// CHECK9-NEXT:    [[TMP107:%.*]] = load i64, i64* [[N_CASTED32]], align 8
-// CHECK9-NEXT:    [[TMP108:%.*]] = load double*, double** [[A]], align 8
-// CHECK9-NEXT:    [[TMP109:%.*]] = load double*, double** [[B]], align 8
-// CHECK9-NEXT:    [[TMP110:%.*]] = load double*, double** [[C]], align 8
-// CHECK9-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP107]], i64* [[TMP112]], align 8
-// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP107]], i64* [[TMP114]], align 8
-// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP115]], align 8
-// CHECK9-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to double**
-// CHECK9-NEXT:    store double* [[TMP108]], double** [[TMP117]], align 8
-// CHECK9-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP119:%.*]] = bitcast i8** [[TMP118]] to double**
-// CHECK9-NEXT:    store double* [[TMP108]], double** [[TMP119]], align 8
-// CHECK9-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP120]], align 8
-// CHECK9-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to double**
-// CHECK9-NEXT:    store double* [[TMP109]], double** [[TMP122]], align 8
-// CHECK9-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to double**
-// CHECK9-NEXT:    store double* [[TMP109]], double** [[TMP124]], align 8
-// CHECK9-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP125]], align 8
-// CHECK9-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to double**
-// CHECK9-NEXT:    store double* [[TMP110]], double** [[TMP127]], align 8
-// CHECK9-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to double**
-// CHECK9-NEXT:    store double* [[TMP110]], double** [[TMP129]], align 8
-// CHECK9-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP130]], align 8
-// CHECK9-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP133:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP133]], i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK9-NEXT:    [[TMP134:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK9-NEXT:    [[SUB40:%.*]] = sub nsw i32 [[TMP134]], 0
-// CHECK9-NEXT:    [[DIV41:%.*]] = sdiv i32 [[SUB40]], 1
-// CHECK9-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[DIV41]], 1
-// CHECK9-NEXT:    store i32 [[SUB42]], i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK9-NEXT:    [[TMP135:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK9-NEXT:    [[ADD43:%.*]] = add nsw i32 [[TMP135]], 1
-// CHECK9-NEXT:    [[TMP136:%.*]] = zext i32 [[ADD43]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP136]])
-// CHECK9-NEXT:    [[TMP137:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l477.region_id, i32 4, i8** [[TMP131]], i8** [[TMP132]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP138:%.*]] = icmp ne i32 [[TMP137]], 0
-// CHECK9-NEXT:    br i1 [[TMP138]], label [[OMP_OFFLOAD_FAILED44:%.*]], label [[OMP_OFFLOAD_CONT45:%.*]]
-// CHECK9:       omp_offload.failed44:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l477(i64 [[TMP107]], double* [[TMP108]], double* [[TMP109]], double* [[TMP110]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT45]]
-// CHECK9:       omp_offload.cont45:
-// CHECK9-NEXT:    [[TMP139:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK9-NEXT:    [[CONV47:%.*]] = bitcast i64* [[CH_CASTED46]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP139]], i32* [[CONV47]], align 4
-// CHECK9-NEXT:    [[TMP140:%.*]] = load i64, i64* [[CH_CASTED46]], align 8
-// CHECK9-NEXT:    [[TMP141:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV49:%.*]] = bitcast i64* [[N_CASTED48]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP141]], i32* [[CONV49]], align 4
-// CHECK9-NEXT:    [[TMP142:%.*]] = load i64, i64* [[N_CASTED48]], align 8
-// CHECK9-NEXT:    [[TMP143:%.*]] = load double*, double** [[A]], align 8
-// CHECK9-NEXT:    [[TMP144:%.*]] = load double*, double** [[B]], align 8
-// CHECK9-NEXT:    [[TMP145:%.*]] = load double*, double** [[C]], align 8
-// CHECK9-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP140]], i64* [[TMP147]], align 8
-// CHECK9-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP140]], i64* [[TMP149]], align 8
-// CHECK9-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP150]], align 8
-// CHECK9-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP142]], i64* [[TMP152]], align 8
-// CHECK9-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP142]], i64* [[TMP154]], align 8
-// CHECK9-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP155]], align 8
-// CHECK9-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to double**
-// CHECK9-NEXT:    store double* [[TMP143]], double** [[TMP157]], align 8
-// CHECK9-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to double**
-// CHECK9-NEXT:    store double* [[TMP143]], double** [[TMP159]], align 8
-// CHECK9-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP160]], align 8
-// CHECK9-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP162:%.*]] = bitcast i8** [[TMP161]] to double**
-// CHECK9-NEXT:    store double* [[TMP144]], double** [[TMP162]], align 8
-// CHECK9-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP164:%.*]] = bitcast i8** [[TMP163]] to double**
-// CHECK9-NEXT:    store double* [[TMP144]], double** [[TMP164]], align 8
-// CHECK9-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP165]], align 8
-// CHECK9-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to double**
-// CHECK9-NEXT:    store double* [[TMP145]], double** [[TMP167]], align 8
-// CHECK9-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to double**
-// CHECK9-NEXT:    store double* [[TMP145]], double** [[TMP169]], align 8
-// CHECK9-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 4
-// CHECK9-NEXT:    store i8* null, i8** [[TMP170]], align 8
-// CHECK9-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP70:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP70]], 1
+// CHECK9-NEXT:    [[TMP71:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP71]])
+// CHECK9-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP72]], align 4
+// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP73]], align 4
+// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP66]], i8*** [[TMP74]], align 8
+// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP67]], i8*** [[TMP75]], align 8
+// CHECK9-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP76]], align 8
+// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP77]], align 8
+// CHECK9-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP78]], align 8
+// CHECK9-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP79]], align 8
+// CHECK9-NEXT:    [[TMP80:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l407.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK9-NEXT:    [[TMP81:%.*]] = icmp ne i32 [[TMP80]], 0
+// CHECK9-NEXT:    br i1 [[TMP81]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK9:       omp_offload.failed16:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l407(i64 [[TMP42]], double* [[TMP43]], double* [[TMP44]], double* [[TMP45]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK9:       omp_offload.cont17:
+// CHECK9-NEXT:    [[TMP82:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK9-NEXT:    [[CONV18:%.*]] = bitcast i64* [[CH_CASTED]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP82]], i32* [[CONV18]], align 4
+// CHECK9-NEXT:    [[TMP83:%.*]] = load i64, i64* [[CH_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP84:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP84]], i32* [[CONV20]], align 4
+// CHECK9-NEXT:    [[TMP85:%.*]] = load i64, i64* [[N_CASTED19]], align 8
+// CHECK9-NEXT:    [[TMP86:%.*]] = load double*, double** [[A]], align 8
+// CHECK9-NEXT:    [[TMP87:%.*]] = load double*, double** [[B]], align 8
+// CHECK9-NEXT:    [[TMP88:%.*]] = load double*, double** [[C]], align 8
+// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP83]], i64* [[TMP90]], align 8
+// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP83]], i64* [[TMP92]], align 8
+// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP93]], align 8
+// CHECK9-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP85]], i64* [[TMP95]], align 8
+// CHECK9-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP85]], i64* [[TMP97]], align 8
+// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP98]], align 8
+// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP100:%.*]] = bitcast i8** [[TMP99]] to double**
+// CHECK9-NEXT:    store double* [[TMP86]], double** [[TMP100]], align 8
+// CHECK9-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to double**
+// CHECK9-NEXT:    store double* [[TMP86]], double** [[TMP102]], align 8
+// CHECK9-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP103]], align 8
+// CHECK9-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to double**
+// CHECK9-NEXT:    store double* [[TMP87]], double** [[TMP105]], align 8
+// CHECK9-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to double**
+// CHECK9-NEXT:    store double* [[TMP87]], double** [[TMP107]], align 8
+// CHECK9-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP108]], align 8
+// CHECK9-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to double**
+// CHECK9-NEXT:    store double* [[TMP88]], double** [[TMP110]], align 8
+// CHECK9-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to double**
+// CHECK9-NEXT:    store double* [[TMP88]], double** [[TMP112]], align 8
+// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 4
+// CHECK9-NEXT:    store i8* null, i8** [[TMP113]], align 8
+// CHECK9-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP116:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP116]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK9-NEXT:    [[TMP117:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK9-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP117]], 0
+// CHECK9-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
+// CHECK9-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
+// CHECK9-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK9-NEXT:    [[TMP118:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK9-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP118]], 1
+// CHECK9-NEXT:    [[TMP119:%.*]] = zext i32 [[ADD30]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP119]])
+// CHECK9-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP120]], align 4
+// CHECK9-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP121]], align 4
+// CHECK9-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP114]], i8*** [[TMP122]], align 8
+// CHECK9-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP115]], i8*** [[TMP123]], align 8
+// CHECK9-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP124]], align 8
+// CHECK9-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP125]], align 8
+// CHECK9-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP126]], align 8
+// CHECK9-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP127]], align 8
+// CHECK9-NEXT:    [[TMP128:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l446.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK9-NEXT:    [[TMP129:%.*]] = icmp ne i32 [[TMP128]], 0
+// CHECK9-NEXT:    br i1 [[TMP129]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK9:       omp_offload.failed32:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l446(i64 [[TMP83]], i64 [[TMP85]], double* [[TMP86]], double* [[TMP87]], double* [[TMP88]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK9:       omp_offload.cont33:
+// CHECK9-NEXT:    [[TMP130:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV35:%.*]] = bitcast i64* [[N_CASTED34]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP130]], i32* [[CONV35]], align 4
+// CHECK9-NEXT:    [[TMP131:%.*]] = load i64, i64* [[N_CASTED34]], align 8
+// CHECK9-NEXT:    [[TMP132:%.*]] = load double*, double** [[A]], align 8
+// CHECK9-NEXT:    [[TMP133:%.*]] = load double*, double** [[B]], align 8
+// CHECK9-NEXT:    [[TMP134:%.*]] = load double*, double** [[C]], align 8
+// CHECK9-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP131]], i64* [[TMP136]], align 8
+// CHECK9-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP131]], i64* [[TMP138]], align 8
+// CHECK9-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP139]], align 8
+// CHECK9-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to double**
+// CHECK9-NEXT:    store double* [[TMP132]], double** [[TMP141]], align 8
+// CHECK9-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to double**
+// CHECK9-NEXT:    store double* [[TMP132]], double** [[TMP143]], align 8
+// CHECK9-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP144]], align 8
+// CHECK9-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to double**
+// CHECK9-NEXT:    store double* [[TMP133]], double** [[TMP146]], align 8
+// CHECK9-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to double**
+// CHECK9-NEXT:    store double* [[TMP133]], double** [[TMP148]], align 8
+// CHECK9-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP149]], align 8
+// CHECK9-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to double**
+// CHECK9-NEXT:    store double* [[TMP134]], double** [[TMP151]], align 8
+// CHECK9-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to double**
+// CHECK9-NEXT:    store double* [[TMP134]], double** [[TMP153]], align 8
+// CHECK9-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP154]], align 8
+// CHECK9-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP157:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP157]], i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK9-NEXT:    [[TMP158:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK9-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[TMP158]], 0
+// CHECK9-NEXT:    [[DIV43:%.*]] = sdiv i32 [[SUB42]], 1
+// CHECK9-NEXT:    [[SUB44:%.*]] = sub nsw i32 [[DIV43]], 1
+// CHECK9-NEXT:    store i32 [[SUB44]], i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK9-NEXT:    [[TMP159:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK9-NEXT:    [[ADD45:%.*]] = add nsw i32 [[TMP159]], 1
+// CHECK9-NEXT:    [[TMP160:%.*]] = zext i32 [[ADD45]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP160]])
+// CHECK9-NEXT:    [[KERNEL_ARGS46:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP161]], align 4
+// CHECK9-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP162]], align 4
+// CHECK9-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP155]], i8*** [[TMP163]], align 8
+// CHECK9-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP156]], i8*** [[TMP164]], align 8
+// CHECK9-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP165]], align 8
+// CHECK9-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP166]], align 8
+// CHECK9-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP167]], align 8
+// CHECK9-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP168]], align 8
+// CHECK9-NEXT:    [[TMP169:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l477.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]])
+// CHECK9-NEXT:    [[TMP170:%.*]] = icmp ne i32 [[TMP169]], 0
+// CHECK9-NEXT:    br i1 [[TMP170]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
+// CHECK9:       omp_offload.failed47:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l477(i64 [[TMP131]], double* [[TMP132]], double* [[TMP133]], double* [[TMP134]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
+// CHECK9:       omp_offload.cont48:
+// CHECK9-NEXT:    [[TMP171:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK9-NEXT:    [[CONV50:%.*]] = bitcast i64* [[CH_CASTED49]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP171]], i32* [[CONV50]], align 4
+// CHECK9-NEXT:    [[TMP172:%.*]] = load i64, i64* [[CH_CASTED49]], align 8
 // CHECK9-NEXT:    [[TMP173:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP173]], i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK9-NEXT:    [[TMP174:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK9-NEXT:    [[SUB56:%.*]] = sub nsw i32 [[TMP174]], 0
-// CHECK9-NEXT:    [[DIV57:%.*]] = sdiv i32 [[SUB56]], 1
-// CHECK9-NEXT:    [[SUB58:%.*]] = sub nsw i32 [[DIV57]], 1
-// CHECK9-NEXT:    store i32 [[SUB58]], i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK9-NEXT:    [[TMP175:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK9-NEXT:    [[ADD59:%.*]] = add nsw i32 [[TMP175]], 1
-// CHECK9-NEXT:    [[TMP176:%.*]] = zext i32 [[ADD59]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP176]])
-// CHECK9-NEXT:    [[TMP177:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l505.region_id, i32 5, i8** [[TMP171]], i8** [[TMP172]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP178:%.*]] = icmp ne i32 [[TMP177]], 0
-// CHECK9-NEXT:    br i1 [[TMP178]], label [[OMP_OFFLOAD_FAILED60:%.*]], label [[OMP_OFFLOAD_CONT61:%.*]]
-// CHECK9:       omp_offload.failed60:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l505(i64 [[TMP140]], i64 [[TMP142]], double* [[TMP143]], double* [[TMP144]], double* [[TMP145]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT61]]
-// CHECK9:       omp_offload.cont61:
-// CHECK9-NEXT:    [[TMP179:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV63:%.*]] = bitcast i64* [[N_CASTED62]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP179]], i32* [[CONV63]], align 4
-// CHECK9-NEXT:    [[TMP180:%.*]] = load i64, i64* [[N_CASTED62]], align 8
-// CHECK9-NEXT:    [[TMP181:%.*]] = load double*, double** [[A]], align 8
-// CHECK9-NEXT:    [[TMP182:%.*]] = load double*, double** [[B]], align 8
-// CHECK9-NEXT:    [[TMP183:%.*]] = load double*, double** [[C]], align 8
-// CHECK9-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP185:%.*]] = bitcast i8** [[TMP184]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP180]], i64* [[TMP185]], align 8
-// CHECK9-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP187:%.*]] = bitcast i8** [[TMP186]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP180]], i64* [[TMP187]], align 8
-// CHECK9-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP188]], align 8
-// CHECK9-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP190:%.*]] = bitcast i8** [[TMP189]] to double**
-// CHECK9-NEXT:    store double* [[TMP181]], double** [[TMP190]], align 8
-// CHECK9-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP192:%.*]] = bitcast i8** [[TMP191]] to double**
-// CHECK9-NEXT:    store double* [[TMP181]], double** [[TMP192]], align 8
-// CHECK9-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP193]], align 8
-// CHECK9-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP195:%.*]] = bitcast i8** [[TMP194]] to double**
-// CHECK9-NEXT:    store double* [[TMP182]], double** [[TMP195]], align 8
-// CHECK9-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP197:%.*]] = bitcast i8** [[TMP196]] to double**
-// CHECK9-NEXT:    store double* [[TMP182]], double** [[TMP197]], align 8
-// CHECK9-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP198]], align 8
-// CHECK9-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP200:%.*]] = bitcast i8** [[TMP199]] to double**
-// CHECK9-NEXT:    store double* [[TMP183]], double** [[TMP200]], align 8
-// CHECK9-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP202:%.*]] = bitcast i8** [[TMP201]] to double**
-// CHECK9-NEXT:    store double* [[TMP183]], double** [[TMP202]], align 8
-// CHECK9-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP203]], align 8
-// CHECK9-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP206:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP206]], i32* [[DOTCAPTURE_EXPR_68]], align 4
-// CHECK9-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_68]], align 4
-// CHECK9-NEXT:    [[SUB70:%.*]] = sub nsw i32 [[TMP207]], 0
-// CHECK9-NEXT:    [[DIV71:%.*]] = sdiv i32 [[SUB70]], 1
-// CHECK9-NEXT:    [[SUB72:%.*]] = sub nsw i32 [[DIV71]], 1
-// CHECK9-NEXT:    store i32 [[SUB72]], i32* [[DOTCAPTURE_EXPR_69]], align 4
-// CHECK9-NEXT:    [[TMP208:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_69]], align 4
-// CHECK9-NEXT:    [[ADD73:%.*]] = add nsw i32 [[TMP208]], 1
-// CHECK9-NEXT:    [[TMP209:%.*]] = zext i32 [[ADD73]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP209]])
-// CHECK9-NEXT:    [[TMP210:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l535.region_id, i32 4, i8** [[TMP204]], i8** [[TMP205]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP211:%.*]] = icmp ne i32 [[TMP210]], 0
-// CHECK9-NEXT:    br i1 [[TMP211]], label [[OMP_OFFLOAD_FAILED74:%.*]], label [[OMP_OFFLOAD_CONT75:%.*]]
-// CHECK9:       omp_offload.failed74:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l535(i64 [[TMP180]], double* [[TMP181]], double* [[TMP182]], double* [[TMP183]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT75]]
-// CHECK9:       omp_offload.cont75:
-// CHECK9-NEXT:    [[TMP212:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK9-NEXT:    [[CONV77:%.*]] = bitcast i64* [[CH_CASTED76]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP212]], i32* [[CONV77]], align 4
-// CHECK9-NEXT:    [[TMP213:%.*]] = load i64, i64* [[CH_CASTED76]], align 8
-// CHECK9-NEXT:    [[TMP214:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV79:%.*]] = bitcast i64* [[N_CASTED78]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP214]], i32* [[CONV79]], align 4
-// CHECK9-NEXT:    [[TMP215:%.*]] = load i64, i64* [[N_CASTED78]], align 8
-// CHECK9-NEXT:    [[TMP216:%.*]] = load double*, double** [[A]], align 8
-// CHECK9-NEXT:    [[TMP217:%.*]] = load double*, double** [[B]], align 8
-// CHECK9-NEXT:    [[TMP218:%.*]] = load double*, double** [[C]], align 8
-// CHECK9-NEXT:    [[TMP219:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP220:%.*]] = bitcast i8** [[TMP219]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP213]], i64* [[TMP220]], align 8
-// CHECK9-NEXT:    [[TMP221:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP222:%.*]] = bitcast i8** [[TMP221]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP213]], i64* [[TMP222]], align 8
-// CHECK9-NEXT:    [[TMP223:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP223]], align 8
-// CHECK9-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 1
+// CHECK9-NEXT:    [[CONV52:%.*]] = bitcast i64* [[N_CASTED51]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP173]], i32* [[CONV52]], align 4
+// CHECK9-NEXT:    [[TMP174:%.*]] = load i64, i64* [[N_CASTED51]], align 8
+// CHECK9-NEXT:    [[TMP175:%.*]] = load double*, double** [[A]], align 8
+// CHECK9-NEXT:    [[TMP176:%.*]] = load double*, double** [[B]], align 8
+// CHECK9-NEXT:    [[TMP177:%.*]] = load double*, double** [[C]], align 8
+// CHECK9-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP172]], i64* [[TMP179]], align 8
+// CHECK9-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP172]], i64* [[TMP181]], align 8
+// CHECK9-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP182]], align 8
+// CHECK9-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP184:%.*]] = bitcast i8** [[TMP183]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP174]], i64* [[TMP184]], align 8
+// CHECK9-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP186:%.*]] = bitcast i8** [[TMP185]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP174]], i64* [[TMP186]], align 8
+// CHECK9-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP187]], align 8
+// CHECK9-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP189:%.*]] = bitcast i8** [[TMP188]] to double**
+// CHECK9-NEXT:    store double* [[TMP175]], double** [[TMP189]], align 8
+// CHECK9-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP191:%.*]] = bitcast i8** [[TMP190]] to double**
+// CHECK9-NEXT:    store double* [[TMP175]], double** [[TMP191]], align 8
+// CHECK9-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP192]], align 8
+// CHECK9-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP194:%.*]] = bitcast i8** [[TMP193]] to double**
+// CHECK9-NEXT:    store double* [[TMP176]], double** [[TMP194]], align 8
+// CHECK9-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP196:%.*]] = bitcast i8** [[TMP195]] to double**
+// CHECK9-NEXT:    store double* [[TMP176]], double** [[TMP196]], align 8
+// CHECK9-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP197]], align 8
+// CHECK9-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP199:%.*]] = bitcast i8** [[TMP198]] to double**
+// CHECK9-NEXT:    store double* [[TMP177]], double** [[TMP199]], align 8
+// CHECK9-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP201:%.*]] = bitcast i8** [[TMP200]] to double**
+// CHECK9-NEXT:    store double* [[TMP177]], double** [[TMP201]], align 8
+// CHECK9-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 4
+// CHECK9-NEXT:    store i8* null, i8** [[TMP202]], align 8
+// CHECK9-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP205:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP205]], i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK9-NEXT:    [[TMP206:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK9-NEXT:    [[SUB59:%.*]] = sub nsw i32 [[TMP206]], 0
+// CHECK9-NEXT:    [[DIV60:%.*]] = sdiv i32 [[SUB59]], 1
+// CHECK9-NEXT:    [[SUB61:%.*]] = sub nsw i32 [[DIV60]], 1
+// CHECK9-NEXT:    store i32 [[SUB61]], i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK9-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK9-NEXT:    [[ADD62:%.*]] = add nsw i32 [[TMP207]], 1
+// CHECK9-NEXT:    [[TMP208:%.*]] = zext i32 [[ADD62]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP208]])
+// CHECK9-NEXT:    [[KERNEL_ARGS63:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP209:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP209]], align 4
+// CHECK9-NEXT:    [[TMP210:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP210]], align 4
+// CHECK9-NEXT:    [[TMP211:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP203]], i8*** [[TMP211]], align 8
+// CHECK9-NEXT:    [[TMP212:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP204]], i8*** [[TMP212]], align 8
+// CHECK9-NEXT:    [[TMP213:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP213]], align 8
+// CHECK9-NEXT:    [[TMP214:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP214]], align 8
+// CHECK9-NEXT:    [[TMP215:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP215]], align 8
+// CHECK9-NEXT:    [[TMP216:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP216]], align 8
+// CHECK9-NEXT:    [[TMP217:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l505.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]])
+// CHECK9-NEXT:    [[TMP218:%.*]] = icmp ne i32 [[TMP217]], 0
+// CHECK9-NEXT:    br i1 [[TMP218]], label [[OMP_OFFLOAD_FAILED64:%.*]], label [[OMP_OFFLOAD_CONT65:%.*]]
+// CHECK9:       omp_offload.failed64:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l505(i64 [[TMP172]], i64 [[TMP174]], double* [[TMP175]], double* [[TMP176]], double* [[TMP177]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT65]]
+// CHECK9:       omp_offload.cont65:
+// CHECK9-NEXT:    [[TMP219:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV67:%.*]] = bitcast i64* [[N_CASTED66]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP219]], i32* [[CONV67]], align 4
+// CHECK9-NEXT:    [[TMP220:%.*]] = load i64, i64* [[N_CASTED66]], align 8
+// CHECK9-NEXT:    [[TMP221:%.*]] = load double*, double** [[A]], align 8
+// CHECK9-NEXT:    [[TMP222:%.*]] = load double*, double** [[B]], align 8
+// CHECK9-NEXT:    [[TMP223:%.*]] = load double*, double** [[C]], align 8
+// CHECK9-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP225:%.*]] = bitcast i8** [[TMP224]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP215]], i64* [[TMP225]], align 8
-// CHECK9-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 1
+// CHECK9-NEXT:    store i64 [[TMP220]], i64* [[TMP225]], align 8
+// CHECK9-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP227:%.*]] = bitcast i8** [[TMP226]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP215]], i64* [[TMP227]], align 8
-// CHECK9-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 1
+// CHECK9-NEXT:    store i64 [[TMP220]], i64* [[TMP227]], align 8
+// CHECK9-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 0
 // CHECK9-NEXT:    store i8* null, i8** [[TMP228]], align 8
-// CHECK9-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 1
 // CHECK9-NEXT:    [[TMP230:%.*]] = bitcast i8** [[TMP229]] to double**
-// CHECK9-NEXT:    store double* [[TMP216]], double** [[TMP230]], align 8
-// CHECK9-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 2
+// CHECK9-NEXT:    store double* [[TMP221]], double** [[TMP230]], align 8
+// CHECK9-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 1
 // CHECK9-NEXT:    [[TMP232:%.*]] = bitcast i8** [[TMP231]] to double**
-// CHECK9-NEXT:    store double* [[TMP216]], double** [[TMP232]], align 8
-// CHECK9-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 2
+// CHECK9-NEXT:    store double* [[TMP221]], double** [[TMP232]], align 8
+// CHECK9-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 1
 // CHECK9-NEXT:    store i8* null, i8** [[TMP233]], align 8
-// CHECK9-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 2
 // CHECK9-NEXT:    [[TMP235:%.*]] = bitcast i8** [[TMP234]] to double**
-// CHECK9-NEXT:    store double* [[TMP217]], double** [[TMP235]], align 8
-// CHECK9-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 3
+// CHECK9-NEXT:    store double* [[TMP222]], double** [[TMP235]], align 8
+// CHECK9-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 2
 // CHECK9-NEXT:    [[TMP237:%.*]] = bitcast i8** [[TMP236]] to double**
-// CHECK9-NEXT:    store double* [[TMP217]], double** [[TMP237]], align 8
-// CHECK9-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 3
+// CHECK9-NEXT:    store double* [[TMP222]], double** [[TMP237]], align 8
+// CHECK9-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 2
 // CHECK9-NEXT:    store i8* null, i8** [[TMP238]], align 8
-// CHECK9-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 3
 // CHECK9-NEXT:    [[TMP240:%.*]] = bitcast i8** [[TMP239]] to double**
-// CHECK9-NEXT:    store double* [[TMP218]], double** [[TMP240]], align 8
-// CHECK9-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 4
+// CHECK9-NEXT:    store double* [[TMP223]], double** [[TMP240]], align 8
+// CHECK9-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 3
 // CHECK9-NEXT:    [[TMP242:%.*]] = bitcast i8** [[TMP241]] to double**
-// CHECK9-NEXT:    store double* [[TMP218]], double** [[TMP242]], align 8
-// CHECK9-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 4
+// CHECK9-NEXT:    store double* [[TMP223]], double** [[TMP242]], align 8
+// CHECK9-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 3
 // CHECK9-NEXT:    store i8* null, i8** [[TMP243]], align 8
-// CHECK9-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP246:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_84]], align 4
-// CHECK9-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_84]], align 4
-// CHECK9-NEXT:    [[SUB86:%.*]] = sub nsw i32 [[TMP247]], 0
-// CHECK9-NEXT:    [[DIV87:%.*]] = sdiv i32 [[SUB86]], 1
-// CHECK9-NEXT:    [[SUB88:%.*]] = sub nsw i32 [[DIV87]], 1
-// CHECK9-NEXT:    store i32 [[SUB88]], i32* [[DOTCAPTURE_EXPR_85]], align 4
-// CHECK9-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_85]], align 4
-// CHECK9-NEXT:    [[ADD89:%.*]] = add nsw i32 [[TMP248]], 1
-// CHECK9-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD89]] to i64
+// CHECK9-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_72]], align 4
+// CHECK9-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_72]], align 4
+// CHECK9-NEXT:    [[SUB74:%.*]] = sub nsw i32 [[TMP247]], 0
+// CHECK9-NEXT:    [[DIV75:%.*]] = sdiv i32 [[SUB74]], 1
+// CHECK9-NEXT:    [[SUB76:%.*]] = sub nsw i32 [[DIV75]], 1
+// CHECK9-NEXT:    store i32 [[SUB76]], i32* [[DOTCAPTURE_EXPR_73]], align 4
+// CHECK9-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_73]], align 4
+// CHECK9-NEXT:    [[ADD77:%.*]] = add nsw i32 [[TMP248]], 1
+// CHECK9-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD77]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP249]])
-// CHECK9-NEXT:    [[TMP250:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l561.region_id, i32 5, i8** [[TMP244]], i8** [[TMP245]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP251:%.*]] = icmp ne i32 [[TMP250]], 0
-// CHECK9-NEXT:    br i1 [[TMP251]], label [[OMP_OFFLOAD_FAILED90:%.*]], label [[OMP_OFFLOAD_CONT91:%.*]]
-// CHECK9:       omp_offload.failed90:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l561(i64 [[TMP213]], i64 [[TMP215]], double* [[TMP216]], double* [[TMP217]], double* [[TMP218]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT91]]
-// CHECK9:       omp_offload.cont91:
+// CHECK9-NEXT:    [[KERNEL_ARGS78:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP250:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP250]], align 4
+// CHECK9-NEXT:    [[TMP251:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP251]], align 4
+// CHECK9-NEXT:    [[TMP252:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP244]], i8*** [[TMP252]], align 8
+// CHECK9-NEXT:    [[TMP253:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP245]], i8*** [[TMP253]], align 8
+// CHECK9-NEXT:    [[TMP254:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP254]], align 8
+// CHECK9-NEXT:    [[TMP255:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP255]], align 8
+// CHECK9-NEXT:    [[TMP256:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP256]], align 8
+// CHECK9-NEXT:    [[TMP257:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP257]], align 8
+// CHECK9-NEXT:    [[TMP258:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l535.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]])
+// CHECK9-NEXT:    [[TMP259:%.*]] = icmp ne i32 [[TMP258]], 0
+// CHECK9-NEXT:    br i1 [[TMP259]], label [[OMP_OFFLOAD_FAILED79:%.*]], label [[OMP_OFFLOAD_CONT80:%.*]]
+// CHECK9:       omp_offload.failed79:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l535(i64 [[TMP220]], double* [[TMP221]], double* [[TMP222]], double* [[TMP223]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT80]]
+// CHECK9:       omp_offload.cont80:
+// CHECK9-NEXT:    [[TMP260:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK9-NEXT:    [[CONV82:%.*]] = bitcast i64* [[CH_CASTED81]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP260]], i32* [[CONV82]], align 4
+// CHECK9-NEXT:    [[TMP261:%.*]] = load i64, i64* [[CH_CASTED81]], align 8
+// CHECK9-NEXT:    [[TMP262:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV84:%.*]] = bitcast i64* [[N_CASTED83]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP262]], i32* [[CONV84]], align 4
+// CHECK9-NEXT:    [[TMP263:%.*]] = load i64, i64* [[N_CASTED83]], align 8
+// CHECK9-NEXT:    [[TMP264:%.*]] = load double*, double** [[A]], align 8
+// CHECK9-NEXT:    [[TMP265:%.*]] = load double*, double** [[B]], align 8
+// CHECK9-NEXT:    [[TMP266:%.*]] = load double*, double** [[C]], align 8
+// CHECK9-NEXT:    [[TMP267:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP268:%.*]] = bitcast i8** [[TMP267]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP261]], i64* [[TMP268]], align 8
+// CHECK9-NEXT:    [[TMP269:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP270:%.*]] = bitcast i8** [[TMP269]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP261]], i64* [[TMP270]], align 8
+// CHECK9-NEXT:    [[TMP271:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP271]], align 8
+// CHECK9-NEXT:    [[TMP272:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP273:%.*]] = bitcast i8** [[TMP272]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP263]], i64* [[TMP273]], align 8
+// CHECK9-NEXT:    [[TMP274:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP275:%.*]] = bitcast i8** [[TMP274]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP263]], i64* [[TMP275]], align 8
+// CHECK9-NEXT:    [[TMP276:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP276]], align 8
+// CHECK9-NEXT:    [[TMP277:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP278:%.*]] = bitcast i8** [[TMP277]] to double**
+// CHECK9-NEXT:    store double* [[TMP264]], double** [[TMP278]], align 8
+// CHECK9-NEXT:    [[TMP279:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP280:%.*]] = bitcast i8** [[TMP279]] to double**
+// CHECK9-NEXT:    store double* [[TMP264]], double** [[TMP280]], align 8
+// CHECK9-NEXT:    [[TMP281:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP281]], align 8
+// CHECK9-NEXT:    [[TMP282:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP283:%.*]] = bitcast i8** [[TMP282]] to double**
+// CHECK9-NEXT:    store double* [[TMP265]], double** [[TMP283]], align 8
+// CHECK9-NEXT:    [[TMP284:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP285:%.*]] = bitcast i8** [[TMP284]] to double**
+// CHECK9-NEXT:    store double* [[TMP265]], double** [[TMP285]], align 8
+// CHECK9-NEXT:    [[TMP286:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP286]], align 8
+// CHECK9-NEXT:    [[TMP287:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP288:%.*]] = bitcast i8** [[TMP287]] to double**
+// CHECK9-NEXT:    store double* [[TMP266]], double** [[TMP288]], align 8
+// CHECK9-NEXT:    [[TMP289:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP290:%.*]] = bitcast i8** [[TMP289]] to double**
+// CHECK9-NEXT:    store double* [[TMP266]], double** [[TMP290]], align 8
+// CHECK9-NEXT:    [[TMP291:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 4
+// CHECK9-NEXT:    store i8* null, i8** [[TMP291]], align 8
+// CHECK9-NEXT:    [[TMP292:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP293:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP294:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP294]], i32* [[DOTCAPTURE_EXPR_89]], align 4
+// CHECK9-NEXT:    [[TMP295:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_89]], align 4
+// CHECK9-NEXT:    [[SUB91:%.*]] = sub nsw i32 [[TMP295]], 0
+// CHECK9-NEXT:    [[DIV92:%.*]] = sdiv i32 [[SUB91]], 1
+// CHECK9-NEXT:    [[SUB93:%.*]] = sub nsw i32 [[DIV92]], 1
+// CHECK9-NEXT:    store i32 [[SUB93]], i32* [[DOTCAPTURE_EXPR_90]], align 4
+// CHECK9-NEXT:    [[TMP296:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_90]], align 4
+// CHECK9-NEXT:    [[ADD94:%.*]] = add nsw i32 [[TMP296]], 1
+// CHECK9-NEXT:    [[TMP297:%.*]] = zext i32 [[ADD94]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP297]])
+// CHECK9-NEXT:    [[KERNEL_ARGS95:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP298:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP298]], align 4
+// CHECK9-NEXT:    [[TMP299:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP299]], align 4
+// CHECK9-NEXT:    [[TMP300:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP292]], i8*** [[TMP300]], align 8
+// CHECK9-NEXT:    [[TMP301:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP293]], i8*** [[TMP301]], align 8
+// CHECK9-NEXT:    [[TMP302:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP302]], align 8
+// CHECK9-NEXT:    [[TMP303:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP303]], align 8
+// CHECK9-NEXT:    [[TMP304:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP304]], align 8
+// CHECK9-NEXT:    [[TMP305:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP305]], align 8
+// CHECK9-NEXT:    [[TMP306:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l561.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]])
+// CHECK9-NEXT:    [[TMP307:%.*]] = icmp ne i32 [[TMP306]], 0
+// CHECK9-NEXT:    br i1 [[TMP307]], label [[OMP_OFFLOAD_FAILED96:%.*]], label [[OMP_OFFLOAD_CONT97:%.*]]
+// CHECK9:       omp_offload.failed96:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l561(i64 [[TMP261]], i64 [[TMP263]], double* [[TMP264]], double* [[TMP265]], double* [[TMP266]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT97]]
+// CHECK9:       omp_offload.cont97:
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    ret i32 [[CALL]]
 //
@@ -7072,43 +7191,43 @@ int main() {
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[CH_CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED18:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED32:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS34:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS35:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS36:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP37:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_38:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_39:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[CH_CASTED46:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED48:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS50:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS51:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS52:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP53:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_54:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_55:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED62:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS64:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS65:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS66:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP67:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_68:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_69:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[CH_CASTED76:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED78:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS80:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS81:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS82:%.*]] = alloca [5 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP83:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_84:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_85:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED34:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS36:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS37:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS38:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP39:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_40:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_41:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[CH_CASTED49:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[N_CASTED51:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS53:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS54:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS55:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP56:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_57:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_58:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED66:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS68:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS69:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS70:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP71:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_72:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_73:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[CH_CASTED81:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[N_CASTED83:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS85:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS86:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS87:%.*]] = alloca [5 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP88:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_89:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_90:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 10000, i32* [[N]], align 4
 // CHECK9-NEXT:    store i32 100, i32* [[CH]], align 4
 // CHECK9-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N]], align 4
@@ -7163,403 +7282,522 @@ int main() {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP29]], 1
 // CHECK9-NEXT:    [[TMP30:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP30]])
-// CHECK9-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l42.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK9-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP32]], align 4
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP25]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP26]], i8*** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l42.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK9-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l42(i64 [[TMP1]], i32* [[TMP2]], i32* [[TMP3]], i32* [[TMP4]]) #[[ATTR2]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N]], align 4
 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP33]], i32* [[CONV4]], align 4
-// CHECK9-NEXT:    [[TMP34:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK9-NEXT:    [[TMP35:%.*]] = load i32*, i32** [[A]], align 8
-// CHECK9-NEXT:    [[TMP36:%.*]] = load i32*, i32** [[B]], align 8
-// CHECK9-NEXT:    [[TMP37:%.*]] = load i32*, i32** [[C]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP41]], align 8
-// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP42]], align 8
-// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP35]], i32** [[TMP44]], align 8
-// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP35]], i32** [[TMP46]], align 8
-// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP47]], align 8
-// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP36]], i32** [[TMP49]], align 8
-// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP36]], i32** [[TMP51]], align 8
-// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP52]], align 8
-// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 3
+// CHECK9-NEXT:    store i32 [[TMP41]], i32* [[CONV4]], align 4
+// CHECK9-NEXT:    [[TMP42:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = load i32*, i32** [[A]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = load i32*, i32** [[B]], align 8
+// CHECK9-NEXT:    [[TMP45:%.*]] = load i32*, i32** [[C]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP49]], align 8
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP43]], i32** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
 // CHECK9-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP37]], i32** [[TMP54]], align 8
-// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP37]], i32** [[TMP56]], align 8
-// CHECK9-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP57]], align 8
-// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP60:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP60]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK9-NEXT:    [[TMP61:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK9-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP61]], 0
+// CHECK9-NEXT:    store i32* [[TMP43]], i32** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP44]], i32** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP44]], i32** [[TMP59]], align 8
+// CHECK9-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP60]], align 8
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP45]], i32** [[TMP62]], align 8
+// CHECK9-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP64:%.*]] = bitcast i8** [[TMP63]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP45]], i32** [[TMP64]], align 8
+// CHECK9-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP65]], align 8
+// CHECK9-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP68]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK9-NEXT:    [[TMP69:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK9-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP69]], 0
 // CHECK9-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK9-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK9-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[TMP62:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP62]], 1
-// CHECK9-NEXT:    [[TMP63:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP63]])
-// CHECK9-NEXT:    [[TMP64:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50.region_id, i32 4, i8** [[TMP58]], i8** [[TMP59]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP65:%.*]] = icmp ne i32 [[TMP64]], 0
-// CHECK9-NEXT:    br i1 [[TMP65]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK9:       omp_offload.failed15:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50(i64 [[TMP34]], i32* [[TMP35]], i32* [[TMP36]], i32* [[TMP37]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK9:       omp_offload.cont16:
-// CHECK9-NEXT:    [[TMP66:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK9-NEXT:    [[CONV17:%.*]] = bitcast i64* [[CH_CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP66]], i32* [[CONV17]], align 4
-// CHECK9-NEXT:    [[TMP67:%.*]] = load i64, i64* [[CH_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV19:%.*]] = bitcast i64* [[N_CASTED18]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP68]], i32* [[CONV19]], align 4
-// CHECK9-NEXT:    [[TMP69:%.*]] = load i64, i64* [[N_CASTED18]], align 8
-// CHECK9-NEXT:    [[TMP70:%.*]] = load i32*, i32** [[A]], align 8
-// CHECK9-NEXT:    [[TMP71:%.*]] = load i32*, i32** [[B]], align 8
-// CHECK9-NEXT:    [[TMP72:%.*]] = load i32*, i32** [[C]], align 8
-// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP67]], i64* [[TMP74]], align 8
-// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP67]], i64* [[TMP76]], align 8
-// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP77]], align 8
-// CHECK9-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP69]], i64* [[TMP79]], align 8
-// CHECK9-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP69]], i64* [[TMP81]], align 8
-// CHECK9-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP82]], align 8
-// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP70]], i32** [[TMP84]], align 8
-// CHECK9-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP70]], i32** [[TMP86]], align 8
-// CHECK9-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP87]], align 8
-// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP71]], i32** [[TMP89]], align 8
-// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP71]], i32** [[TMP91]], align 8
-// CHECK9-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP92]], align 8
-// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP72]], i32** [[TMP94]], align 8
-// CHECK9-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP72]], i32** [[TMP96]], align 8
-// CHECK9-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 4
-// CHECK9-NEXT:    store i8* null, i8** [[TMP97]], align 8
-// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP100]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK9-NEXT:    [[TMP101:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK9-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP101]], 0
-// CHECK9-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
-// CHECK9-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
-// CHECK9-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK9-NEXT:    [[TMP102:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK9-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP102]], 1
-// CHECK9-NEXT:    [[TMP103:%.*]] = zext i32 [[ADD29]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP103]])
-// CHECK9-NEXT:    [[TMP104:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l58.region_id, i32 5, i8** [[TMP98]], i8** [[TMP99]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP105:%.*]] = icmp ne i32 [[TMP104]], 0
-// CHECK9-NEXT:    br i1 [[TMP105]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK9:       omp_offload.failed30:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l58(i64 [[TMP67]], i64 [[TMP69]], i32* [[TMP70]], i32* [[TMP71]], i32* [[TMP72]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK9:       omp_offload.cont31:
-// CHECK9-NEXT:    [[TMP106:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV33:%.*]] = bitcast i64* [[N_CASTED32]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP106]], i32* [[CONV33]], align 4
-// CHECK9-NEXT:    [[TMP107:%.*]] = load i64, i64* [[N_CASTED32]], align 8
-// CHECK9-NEXT:    [[TMP108:%.*]] = load i32*, i32** [[A]], align 8
-// CHECK9-NEXT:    [[TMP109:%.*]] = load i32*, i32** [[B]], align 8
-// CHECK9-NEXT:    [[TMP110:%.*]] = load i32*, i32** [[C]], align 8
-// CHECK9-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP107]], i64* [[TMP112]], align 8
-// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP107]], i64* [[TMP114]], align 8
-// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP115]], align 8
-// CHECK9-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP108]], i32** [[TMP117]], align 8
-// CHECK9-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP119:%.*]] = bitcast i8** [[TMP118]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP108]], i32** [[TMP119]], align 8
-// CHECK9-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP120]], align 8
-// CHECK9-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP109]], i32** [[TMP122]], align 8
-// CHECK9-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP109]], i32** [[TMP124]], align 8
-// CHECK9-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP125]], align 8
-// CHECK9-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP110]], i32** [[TMP127]], align 8
-// CHECK9-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP110]], i32** [[TMP129]], align 8
-// CHECK9-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP130]], align 8
-// CHECK9-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP133:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP133]], i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK9-NEXT:    [[TMP134:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK9-NEXT:    [[SUB40:%.*]] = sub nsw i32 [[TMP134]], 0
-// CHECK9-NEXT:    [[DIV41:%.*]] = sdiv i32 [[SUB40]], 1
-// CHECK9-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[DIV41]], 1
-// CHECK9-NEXT:    store i32 [[SUB42]], i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK9-NEXT:    [[TMP135:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK9-NEXT:    [[ADD43:%.*]] = add nsw i32 [[TMP135]], 1
-// CHECK9-NEXT:    [[TMP136:%.*]] = zext i32 [[ADD43]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP136]])
-// CHECK9-NEXT:    [[TMP137:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l66.region_id, i32 4, i8** [[TMP131]], i8** [[TMP132]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.40, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.41, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP138:%.*]] = icmp ne i32 [[TMP137]], 0
-// CHECK9-NEXT:    br i1 [[TMP138]], label [[OMP_OFFLOAD_FAILED44:%.*]], label [[OMP_OFFLOAD_CONT45:%.*]]
-// CHECK9:       omp_offload.failed44:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l66(i64 [[TMP107]], i32* [[TMP108]], i32* [[TMP109]], i32* [[TMP110]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT45]]
-// CHECK9:       omp_offload.cont45:
-// CHECK9-NEXT:    [[TMP139:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK9-NEXT:    [[CONV47:%.*]] = bitcast i64* [[CH_CASTED46]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP139]], i32* [[CONV47]], align 4
-// CHECK9-NEXT:    [[TMP140:%.*]] = load i64, i64* [[CH_CASTED46]], align 8
-// CHECK9-NEXT:    [[TMP141:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV49:%.*]] = bitcast i64* [[N_CASTED48]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP141]], i32* [[CONV49]], align 4
-// CHECK9-NEXT:    [[TMP142:%.*]] = load i64, i64* [[N_CASTED48]], align 8
-// CHECK9-NEXT:    [[TMP143:%.*]] = load i32*, i32** [[A]], align 8
-// CHECK9-NEXT:    [[TMP144:%.*]] = load i32*, i32** [[B]], align 8
-// CHECK9-NEXT:    [[TMP145:%.*]] = load i32*, i32** [[C]], align 8
-// CHECK9-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP140]], i64* [[TMP147]], align 8
-// CHECK9-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP140]], i64* [[TMP149]], align 8
-// CHECK9-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP150]], align 8
-// CHECK9-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP142]], i64* [[TMP152]], align 8
-// CHECK9-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP142]], i64* [[TMP154]], align 8
-// CHECK9-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP155]], align 8
-// CHECK9-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP143]], i32** [[TMP157]], align 8
-// CHECK9-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP143]], i32** [[TMP159]], align 8
-// CHECK9-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP160]], align 8
-// CHECK9-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP162:%.*]] = bitcast i8** [[TMP161]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP144]], i32** [[TMP162]], align 8
-// CHECK9-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP164:%.*]] = bitcast i8** [[TMP163]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP144]], i32** [[TMP164]], align 8
-// CHECK9-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP165]], align 8
-// CHECK9-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP145]], i32** [[TMP167]], align 8
-// CHECK9-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 4
-// CHECK9-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP145]], i32** [[TMP169]], align 8
-// CHECK9-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i64 0, i64 4
-// CHECK9-NEXT:    store i8* null, i8** [[TMP170]], align 8
-// CHECK9-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP70:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP70]], 1
+// CHECK9-NEXT:    [[TMP71:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP71]])
+// CHECK9-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP72]], align 4
+// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP73]], align 4
+// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP66]], i8*** [[TMP74]], align 8
+// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP67]], i8*** [[TMP75]], align 8
+// CHECK9-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP76]], align 8
+// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP77]], align 8
+// CHECK9-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP78]], align 8
+// CHECK9-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP79]], align 8
+// CHECK9-NEXT:    [[TMP80:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK9-NEXT:    [[TMP81:%.*]] = icmp ne i32 [[TMP80]], 0
+// CHECK9-NEXT:    br i1 [[TMP81]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK9:       omp_offload.failed16:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50(i64 [[TMP42]], i32* [[TMP43]], i32* [[TMP44]], i32* [[TMP45]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK9:       omp_offload.cont17:
+// CHECK9-NEXT:    [[TMP82:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK9-NEXT:    [[CONV18:%.*]] = bitcast i64* [[CH_CASTED]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP82]], i32* [[CONV18]], align 4
+// CHECK9-NEXT:    [[TMP83:%.*]] = load i64, i64* [[CH_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP84:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP84]], i32* [[CONV20]], align 4
+// CHECK9-NEXT:    [[TMP85:%.*]] = load i64, i64* [[N_CASTED19]], align 8
+// CHECK9-NEXT:    [[TMP86:%.*]] = load i32*, i32** [[A]], align 8
+// CHECK9-NEXT:    [[TMP87:%.*]] = load i32*, i32** [[B]], align 8
+// CHECK9-NEXT:    [[TMP88:%.*]] = load i32*, i32** [[C]], align 8
+// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP83]], i64* [[TMP90]], align 8
+// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP83]], i64* [[TMP92]], align 8
+// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP93]], align 8
+// CHECK9-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP85]], i64* [[TMP95]], align 8
+// CHECK9-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP85]], i64* [[TMP97]], align 8
+// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP98]], align 8
+// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP100:%.*]] = bitcast i8** [[TMP99]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP86]], i32** [[TMP100]], align 8
+// CHECK9-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP86]], i32** [[TMP102]], align 8
+// CHECK9-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP103]], align 8
+// CHECK9-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP87]], i32** [[TMP105]], align 8
+// CHECK9-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP87]], i32** [[TMP107]], align 8
+// CHECK9-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP108]], align 8
+// CHECK9-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP88]], i32** [[TMP110]], align 8
+// CHECK9-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP88]], i32** [[TMP112]], align 8
+// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 4
+// CHECK9-NEXT:    store i8* null, i8** [[TMP113]], align 8
+// CHECK9-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP116:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP116]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK9-NEXT:    [[TMP117:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK9-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP117]], 0
+// CHECK9-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
+// CHECK9-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
+// CHECK9-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK9-NEXT:    [[TMP118:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK9-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP118]], 1
+// CHECK9-NEXT:    [[TMP119:%.*]] = zext i32 [[ADD30]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP119]])
+// CHECK9-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP120]], align 4
+// CHECK9-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP121]], align 4
+// CHECK9-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP114]], i8*** [[TMP122]], align 8
+// CHECK9-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP115]], i8*** [[TMP123]], align 8
+// CHECK9-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP124]], align 8
+// CHECK9-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP125]], align 8
+// CHECK9-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP126]], align 8
+// CHECK9-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP127]], align 8
+// CHECK9-NEXT:    [[TMP128:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l58.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK9-NEXT:    [[TMP129:%.*]] = icmp ne i32 [[TMP128]], 0
+// CHECK9-NEXT:    br i1 [[TMP129]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK9:       omp_offload.failed32:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l58(i64 [[TMP83]], i64 [[TMP85]], i32* [[TMP86]], i32* [[TMP87]], i32* [[TMP88]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK9:       omp_offload.cont33:
+// CHECK9-NEXT:    [[TMP130:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV35:%.*]] = bitcast i64* [[N_CASTED34]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP130]], i32* [[CONV35]], align 4
+// CHECK9-NEXT:    [[TMP131:%.*]] = load i64, i64* [[N_CASTED34]], align 8
+// CHECK9-NEXT:    [[TMP132:%.*]] = load i32*, i32** [[A]], align 8
+// CHECK9-NEXT:    [[TMP133:%.*]] = load i32*, i32** [[B]], align 8
+// CHECK9-NEXT:    [[TMP134:%.*]] = load i32*, i32** [[C]], align 8
+// CHECK9-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP131]], i64* [[TMP136]], align 8
+// CHECK9-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP131]], i64* [[TMP138]], align 8
+// CHECK9-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP139]], align 8
+// CHECK9-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP132]], i32** [[TMP141]], align 8
+// CHECK9-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP132]], i32** [[TMP143]], align 8
+// CHECK9-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP144]], align 8
+// CHECK9-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP133]], i32** [[TMP146]], align 8
+// CHECK9-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP133]], i32** [[TMP148]], align 8
+// CHECK9-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP149]], align 8
+// CHECK9-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP134]], i32** [[TMP151]], align 8
+// CHECK9-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP134]], i32** [[TMP153]], align 8
+// CHECK9-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP154]], align 8
+// CHECK9-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP157:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP157]], i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK9-NEXT:    [[TMP158:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK9-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[TMP158]], 0
+// CHECK9-NEXT:    [[DIV43:%.*]] = sdiv i32 [[SUB42]], 1
+// CHECK9-NEXT:    [[SUB44:%.*]] = sub nsw i32 [[DIV43]], 1
+// CHECK9-NEXT:    store i32 [[SUB44]], i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK9-NEXT:    [[TMP159:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK9-NEXT:    [[ADD45:%.*]] = add nsw i32 [[TMP159]], 1
+// CHECK9-NEXT:    [[TMP160:%.*]] = zext i32 [[ADD45]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP160]])
+// CHECK9-NEXT:    [[KERNEL_ARGS46:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP161]], align 4
+// CHECK9-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP162]], align 4
+// CHECK9-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP155]], i8*** [[TMP163]], align 8
+// CHECK9-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP156]], i8*** [[TMP164]], align 8
+// CHECK9-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.40, i32 0, i32 0), i64** [[TMP165]], align 8
+// CHECK9-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.41, i32 0, i32 0), i64** [[TMP166]], align 8
+// CHECK9-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP167]], align 8
+// CHECK9-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP168]], align 8
+// CHECK9-NEXT:    [[TMP169:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l66.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]])
+// CHECK9-NEXT:    [[TMP170:%.*]] = icmp ne i32 [[TMP169]], 0
+// CHECK9-NEXT:    br i1 [[TMP170]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
+// CHECK9:       omp_offload.failed47:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l66(i64 [[TMP131]], i32* [[TMP132]], i32* [[TMP133]], i32* [[TMP134]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
+// CHECK9:       omp_offload.cont48:
+// CHECK9-NEXT:    [[TMP171:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK9-NEXT:    [[CONV50:%.*]] = bitcast i64* [[CH_CASTED49]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP171]], i32* [[CONV50]], align 4
+// CHECK9-NEXT:    [[TMP172:%.*]] = load i64, i64* [[CH_CASTED49]], align 8
 // CHECK9-NEXT:    [[TMP173:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP173]], i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK9-NEXT:    [[TMP174:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK9-NEXT:    [[SUB56:%.*]] = sub nsw i32 [[TMP174]], 0
-// CHECK9-NEXT:    [[DIV57:%.*]] = sdiv i32 [[SUB56]], 1
-// CHECK9-NEXT:    [[SUB58:%.*]] = sub nsw i32 [[DIV57]], 1
-// CHECK9-NEXT:    store i32 [[SUB58]], i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK9-NEXT:    [[TMP175:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK9-NEXT:    [[ADD59:%.*]] = add nsw i32 [[TMP175]], 1
-// CHECK9-NEXT:    [[TMP176:%.*]] = zext i32 [[ADD59]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP176]])
-// CHECK9-NEXT:    [[TMP177:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l74.region_id, i32 5, i8** [[TMP171]], i8** [[TMP172]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.44, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.45, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP178:%.*]] = icmp ne i32 [[TMP177]], 0
-// CHECK9-NEXT:    br i1 [[TMP178]], label [[OMP_OFFLOAD_FAILED60:%.*]], label [[OMP_OFFLOAD_CONT61:%.*]]
-// CHECK9:       omp_offload.failed60:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l74(i64 [[TMP140]], i64 [[TMP142]], i32* [[TMP143]], i32* [[TMP144]], i32* [[TMP145]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT61]]
-// CHECK9:       omp_offload.cont61:
-// CHECK9-NEXT:    [[TMP179:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV63:%.*]] = bitcast i64* [[N_CASTED62]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP179]], i32* [[CONV63]], align 4
-// CHECK9-NEXT:    [[TMP180:%.*]] = load i64, i64* [[N_CASTED62]], align 8
-// CHECK9-NEXT:    [[TMP181:%.*]] = load i32*, i32** [[A]], align 8
-// CHECK9-NEXT:    [[TMP182:%.*]] = load i32*, i32** [[B]], align 8
-// CHECK9-NEXT:    [[TMP183:%.*]] = load i32*, i32** [[C]], align 8
-// CHECK9-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP185:%.*]] = bitcast i8** [[TMP184]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP180]], i64* [[TMP185]], align 8
-// CHECK9-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP187:%.*]] = bitcast i8** [[TMP186]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP180]], i64* [[TMP187]], align 8
-// CHECK9-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP188]], align 8
-// CHECK9-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP190:%.*]] = bitcast i8** [[TMP189]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP181]], i32** [[TMP190]], align 8
-// CHECK9-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP192:%.*]] = bitcast i8** [[TMP191]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP181]], i32** [[TMP192]], align 8
-// CHECK9-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP193]], align 8
-// CHECK9-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP195:%.*]] = bitcast i8** [[TMP194]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP182]], i32** [[TMP195]], align 8
-// CHECK9-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP197:%.*]] = bitcast i8** [[TMP196]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP182]], i32** [[TMP197]], align 8
-// CHECK9-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP198]], align 8
-// CHECK9-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP200:%.*]] = bitcast i8** [[TMP199]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP183]], i32** [[TMP200]], align 8
-// CHECK9-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP202:%.*]] = bitcast i8** [[TMP201]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP183]], i32** [[TMP202]], align 8
-// CHECK9-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS66]], i64 0, i64 3
-// CHECK9-NEXT:    store i8* null, i8** [[TMP203]], align 8
-// CHECK9-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS64]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS65]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP206:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP206]], i32* [[DOTCAPTURE_EXPR_68]], align 4
-// CHECK9-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_68]], align 4
-// CHECK9-NEXT:    [[SUB70:%.*]] = sub nsw i32 [[TMP207]], 0
-// CHECK9-NEXT:    [[DIV71:%.*]] = sdiv i32 [[SUB70]], 1
-// CHECK9-NEXT:    [[SUB72:%.*]] = sub nsw i32 [[DIV71]], 1
-// CHECK9-NEXT:    store i32 [[SUB72]], i32* [[DOTCAPTURE_EXPR_69]], align 4
-// CHECK9-NEXT:    [[TMP208:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_69]], align 4
-// CHECK9-NEXT:    [[ADD73:%.*]] = add nsw i32 [[TMP208]], 1
-// CHECK9-NEXT:    [[TMP209:%.*]] = zext i32 [[ADD73]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP209]])
-// CHECK9-NEXT:    [[TMP210:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l82.region_id, i32 4, i8** [[TMP204]], i8** [[TMP205]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.48, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.49, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP211:%.*]] = icmp ne i32 [[TMP210]], 0
-// CHECK9-NEXT:    br i1 [[TMP211]], label [[OMP_OFFLOAD_FAILED74:%.*]], label [[OMP_OFFLOAD_CONT75:%.*]]
-// CHECK9:       omp_offload.failed74:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l82(i64 [[TMP180]], i32* [[TMP181]], i32* [[TMP182]], i32* [[TMP183]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT75]]
-// CHECK9:       omp_offload.cont75:
-// CHECK9-NEXT:    [[TMP212:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK9-NEXT:    [[CONV77:%.*]] = bitcast i64* [[CH_CASTED76]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP212]], i32* [[CONV77]], align 4
-// CHECK9-NEXT:    [[TMP213:%.*]] = load i64, i64* [[CH_CASTED76]], align 8
-// CHECK9-NEXT:    [[TMP214:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV79:%.*]] = bitcast i64* [[N_CASTED78]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP214]], i32* [[CONV79]], align 4
-// CHECK9-NEXT:    [[TMP215:%.*]] = load i64, i64* [[N_CASTED78]], align 8
-// CHECK9-NEXT:    [[TMP216:%.*]] = load i32*, i32** [[A]], align 8
-// CHECK9-NEXT:    [[TMP217:%.*]] = load i32*, i32** [[B]], align 8
-// CHECK9-NEXT:    [[TMP218:%.*]] = load i32*, i32** [[C]], align 8
-// CHECK9-NEXT:    [[TMP219:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP220:%.*]] = bitcast i8** [[TMP219]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP213]], i64* [[TMP220]], align 8
-// CHECK9-NEXT:    [[TMP221:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP222:%.*]] = bitcast i8** [[TMP221]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP213]], i64* [[TMP222]], align 8
-// CHECK9-NEXT:    [[TMP223:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP223]], align 8
-// CHECK9-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 1
+// CHECK9-NEXT:    [[CONV52:%.*]] = bitcast i64* [[N_CASTED51]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP173]], i32* [[CONV52]], align 4
+// CHECK9-NEXT:    [[TMP174:%.*]] = load i64, i64* [[N_CASTED51]], align 8
+// CHECK9-NEXT:    [[TMP175:%.*]] = load i32*, i32** [[A]], align 8
+// CHECK9-NEXT:    [[TMP176:%.*]] = load i32*, i32** [[B]], align 8
+// CHECK9-NEXT:    [[TMP177:%.*]] = load i32*, i32** [[C]], align 8
+// CHECK9-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP172]], i64* [[TMP179]], align 8
+// CHECK9-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP172]], i64* [[TMP181]], align 8
+// CHECK9-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP182]], align 8
+// CHECK9-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP184:%.*]] = bitcast i8** [[TMP183]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP174]], i64* [[TMP184]], align 8
+// CHECK9-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP186:%.*]] = bitcast i8** [[TMP185]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP174]], i64* [[TMP186]], align 8
+// CHECK9-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP187]], align 8
+// CHECK9-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP189:%.*]] = bitcast i8** [[TMP188]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP175]], i32** [[TMP189]], align 8
+// CHECK9-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP191:%.*]] = bitcast i8** [[TMP190]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP175]], i32** [[TMP191]], align 8
+// CHECK9-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP192]], align 8
+// CHECK9-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP194:%.*]] = bitcast i8** [[TMP193]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP176]], i32** [[TMP194]], align 8
+// CHECK9-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP196:%.*]] = bitcast i8** [[TMP195]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP176]], i32** [[TMP196]], align 8
+// CHECK9-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP197]], align 8
+// CHECK9-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP199:%.*]] = bitcast i8** [[TMP198]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP177]], i32** [[TMP199]], align 8
+// CHECK9-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP201:%.*]] = bitcast i8** [[TMP200]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP177]], i32** [[TMP201]], align 8
+// CHECK9-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 4
+// CHECK9-NEXT:    store i8* null, i8** [[TMP202]], align 8
+// CHECK9-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP205:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP205]], i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK9-NEXT:    [[TMP206:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK9-NEXT:    [[SUB59:%.*]] = sub nsw i32 [[TMP206]], 0
+// CHECK9-NEXT:    [[DIV60:%.*]] = sdiv i32 [[SUB59]], 1
+// CHECK9-NEXT:    [[SUB61:%.*]] = sub nsw i32 [[DIV60]], 1
+// CHECK9-NEXT:    store i32 [[SUB61]], i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK9-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK9-NEXT:    [[ADD62:%.*]] = add nsw i32 [[TMP207]], 1
+// CHECK9-NEXT:    [[TMP208:%.*]] = zext i32 [[ADD62]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP208]])
+// CHECK9-NEXT:    [[KERNEL_ARGS63:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP209:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP209]], align 4
+// CHECK9-NEXT:    [[TMP210:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP210]], align 4
+// CHECK9-NEXT:    [[TMP211:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP203]], i8*** [[TMP211]], align 8
+// CHECK9-NEXT:    [[TMP212:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP204]], i8*** [[TMP212]], align 8
+// CHECK9-NEXT:    [[TMP213:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.44, i32 0, i32 0), i64** [[TMP213]], align 8
+// CHECK9-NEXT:    [[TMP214:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.45, i32 0, i32 0), i64** [[TMP214]], align 8
+// CHECK9-NEXT:    [[TMP215:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP215]], align 8
+// CHECK9-NEXT:    [[TMP216:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP216]], align 8
+// CHECK9-NEXT:    [[TMP217:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l74.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]])
+// CHECK9-NEXT:    [[TMP218:%.*]] = icmp ne i32 [[TMP217]], 0
+// CHECK9-NEXT:    br i1 [[TMP218]], label [[OMP_OFFLOAD_FAILED64:%.*]], label [[OMP_OFFLOAD_CONT65:%.*]]
+// CHECK9:       omp_offload.failed64:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l74(i64 [[TMP172]], i64 [[TMP174]], i32* [[TMP175]], i32* [[TMP176]], i32* [[TMP177]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT65]]
+// CHECK9:       omp_offload.cont65:
+// CHECK9-NEXT:    [[TMP219:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV67:%.*]] = bitcast i64* [[N_CASTED66]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP219]], i32* [[CONV67]], align 4
+// CHECK9-NEXT:    [[TMP220:%.*]] = load i64, i64* [[N_CASTED66]], align 8
+// CHECK9-NEXT:    [[TMP221:%.*]] = load i32*, i32** [[A]], align 8
+// CHECK9-NEXT:    [[TMP222:%.*]] = load i32*, i32** [[B]], align 8
+// CHECK9-NEXT:    [[TMP223:%.*]] = load i32*, i32** [[C]], align 8
+// CHECK9-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP225:%.*]] = bitcast i8** [[TMP224]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP215]], i64* [[TMP225]], align 8
-// CHECK9-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 1
+// CHECK9-NEXT:    store i64 [[TMP220]], i64* [[TMP225]], align 8
+// CHECK9-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP227:%.*]] = bitcast i8** [[TMP226]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP215]], i64* [[TMP227]], align 8
-// CHECK9-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 1
+// CHECK9-NEXT:    store i64 [[TMP220]], i64* [[TMP227]], align 8
+// CHECK9-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 0
 // CHECK9-NEXT:    store i8* null, i8** [[TMP228]], align 8
-// CHECK9-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 1
 // CHECK9-NEXT:    [[TMP230:%.*]] = bitcast i8** [[TMP229]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP216]], i32** [[TMP230]], align 8
-// CHECK9-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 2
+// CHECK9-NEXT:    store i32* [[TMP221]], i32** [[TMP230]], align 8
+// CHECK9-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 1
 // CHECK9-NEXT:    [[TMP232:%.*]] = bitcast i8** [[TMP231]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP216]], i32** [[TMP232]], align 8
-// CHECK9-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 2
+// CHECK9-NEXT:    store i32* [[TMP221]], i32** [[TMP232]], align 8
+// CHECK9-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 1
 // CHECK9-NEXT:    store i8* null, i8** [[TMP233]], align 8
-// CHECK9-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 2
 // CHECK9-NEXT:    [[TMP235:%.*]] = bitcast i8** [[TMP234]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP217]], i32** [[TMP235]], align 8
-// CHECK9-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 3
+// CHECK9-NEXT:    store i32* [[TMP222]], i32** [[TMP235]], align 8
+// CHECK9-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 2
 // CHECK9-NEXT:    [[TMP237:%.*]] = bitcast i8** [[TMP236]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP217]], i32** [[TMP237]], align 8
-// CHECK9-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 3
+// CHECK9-NEXT:    store i32* [[TMP222]], i32** [[TMP237]], align 8
+// CHECK9-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 2
 // CHECK9-NEXT:    store i8* null, i8** [[TMP238]], align 8
-// CHECK9-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 3
 // CHECK9-NEXT:    [[TMP240:%.*]] = bitcast i8** [[TMP239]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP218]], i32** [[TMP240]], align 8
-// CHECK9-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 4
+// CHECK9-NEXT:    store i32* [[TMP223]], i32** [[TMP240]], align 8
+// CHECK9-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 3
 // CHECK9-NEXT:    [[TMP242:%.*]] = bitcast i8** [[TMP241]] to i32**
-// CHECK9-NEXT:    store i32* [[TMP218]], i32** [[TMP242]], align 8
-// CHECK9-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS82]], i64 0, i64 4
+// CHECK9-NEXT:    store i32* [[TMP223]], i32** [[TMP242]], align 8
+// CHECK9-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS70]], i64 0, i64 3
 // CHECK9-NEXT:    store i8* null, i8** [[TMP243]], align 8
-// CHECK9-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS80]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS81]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS68]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS69]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP246:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_84]], align 4
-// CHECK9-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_84]], align 4
-// CHECK9-NEXT:    [[SUB86:%.*]] = sub nsw i32 [[TMP247]], 0
-// CHECK9-NEXT:    [[DIV87:%.*]] = sdiv i32 [[SUB86]], 1
-// CHECK9-NEXT:    [[SUB88:%.*]] = sub nsw i32 [[DIV87]], 1
-// CHECK9-NEXT:    store i32 [[SUB88]], i32* [[DOTCAPTURE_EXPR_85]], align 4
-// CHECK9-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_85]], align 4
-// CHECK9-NEXT:    [[ADD89:%.*]] = add nsw i32 [[TMP248]], 1
-// CHECK9-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD89]] to i64
+// CHECK9-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_72]], align 4
+// CHECK9-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_72]], align 4
+// CHECK9-NEXT:    [[SUB74:%.*]] = sub nsw i32 [[TMP247]], 0
+// CHECK9-NEXT:    [[DIV75:%.*]] = sdiv i32 [[SUB74]], 1
+// CHECK9-NEXT:    [[SUB76:%.*]] = sub nsw i32 [[DIV75]], 1
+// CHECK9-NEXT:    store i32 [[SUB76]], i32* [[DOTCAPTURE_EXPR_73]], align 4
+// CHECK9-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_73]], align 4
+// CHECK9-NEXT:    [[ADD77:%.*]] = add nsw i32 [[TMP248]], 1
+// CHECK9-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD77]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP249]])
-// CHECK9-NEXT:    [[TMP250:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l90.region_id, i32 5, i8** [[TMP244]], i8** [[TMP245]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.52, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.53, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP251:%.*]] = icmp ne i32 [[TMP250]], 0
-// CHECK9-NEXT:    br i1 [[TMP251]], label [[OMP_OFFLOAD_FAILED90:%.*]], label [[OMP_OFFLOAD_CONT91:%.*]]
-// CHECK9:       omp_offload.failed90:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l90(i64 [[TMP213]], i64 [[TMP215]], i32* [[TMP216]], i32* [[TMP217]], i32* [[TMP218]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT91]]
-// CHECK9:       omp_offload.cont91:
+// CHECK9-NEXT:    [[KERNEL_ARGS78:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP250:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP250]], align 4
+// CHECK9-NEXT:    [[TMP251:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP251]], align 4
+// CHECK9-NEXT:    [[TMP252:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP244]], i8*** [[TMP252]], align 8
+// CHECK9-NEXT:    [[TMP253:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP245]], i8*** [[TMP253]], align 8
+// CHECK9-NEXT:    [[TMP254:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.48, i32 0, i32 0), i64** [[TMP254]], align 8
+// CHECK9-NEXT:    [[TMP255:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.49, i32 0, i32 0), i64** [[TMP255]], align 8
+// CHECK9-NEXT:    [[TMP256:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP256]], align 8
+// CHECK9-NEXT:    [[TMP257:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP257]], align 8
+// CHECK9-NEXT:    [[TMP258:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l82.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS78]])
+// CHECK9-NEXT:    [[TMP259:%.*]] = icmp ne i32 [[TMP258]], 0
+// CHECK9-NEXT:    br i1 [[TMP259]], label [[OMP_OFFLOAD_FAILED79:%.*]], label [[OMP_OFFLOAD_CONT80:%.*]]
+// CHECK9:       omp_offload.failed79:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l82(i64 [[TMP220]], i32* [[TMP221]], i32* [[TMP222]], i32* [[TMP223]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT80]]
+// CHECK9:       omp_offload.cont80:
+// CHECK9-NEXT:    [[TMP260:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK9-NEXT:    [[CONV82:%.*]] = bitcast i64* [[CH_CASTED81]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP260]], i32* [[CONV82]], align 4
+// CHECK9-NEXT:    [[TMP261:%.*]] = load i64, i64* [[CH_CASTED81]], align 8
+// CHECK9-NEXT:    [[TMP262:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV84:%.*]] = bitcast i64* [[N_CASTED83]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP262]], i32* [[CONV84]], align 4
+// CHECK9-NEXT:    [[TMP263:%.*]] = load i64, i64* [[N_CASTED83]], align 8
+// CHECK9-NEXT:    [[TMP264:%.*]] = load i32*, i32** [[A]], align 8
+// CHECK9-NEXT:    [[TMP265:%.*]] = load i32*, i32** [[B]], align 8
+// CHECK9-NEXT:    [[TMP266:%.*]] = load i32*, i32** [[C]], align 8
+// CHECK9-NEXT:    [[TMP267:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP268:%.*]] = bitcast i8** [[TMP267]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP261]], i64* [[TMP268]], align 8
+// CHECK9-NEXT:    [[TMP269:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP270:%.*]] = bitcast i8** [[TMP269]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP261]], i64* [[TMP270]], align 8
+// CHECK9-NEXT:    [[TMP271:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP271]], align 8
+// CHECK9-NEXT:    [[TMP272:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP273:%.*]] = bitcast i8** [[TMP272]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP263]], i64* [[TMP273]], align 8
+// CHECK9-NEXT:    [[TMP274:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP275:%.*]] = bitcast i8** [[TMP274]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP263]], i64* [[TMP275]], align 8
+// CHECK9-NEXT:    [[TMP276:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP276]], align 8
+// CHECK9-NEXT:    [[TMP277:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP278:%.*]] = bitcast i8** [[TMP277]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP264]], i32** [[TMP278]], align 8
+// CHECK9-NEXT:    [[TMP279:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP280:%.*]] = bitcast i8** [[TMP279]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP264]], i32** [[TMP280]], align 8
+// CHECK9-NEXT:    [[TMP281:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP281]], align 8
+// CHECK9-NEXT:    [[TMP282:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP283:%.*]] = bitcast i8** [[TMP282]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP265]], i32** [[TMP283]], align 8
+// CHECK9-NEXT:    [[TMP284:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP285:%.*]] = bitcast i8** [[TMP284]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP265]], i32** [[TMP285]], align 8
+// CHECK9-NEXT:    [[TMP286:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP286]], align 8
+// CHECK9-NEXT:    [[TMP287:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP288:%.*]] = bitcast i8** [[TMP287]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP266]], i32** [[TMP288]], align 8
+// CHECK9-NEXT:    [[TMP289:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 4
+// CHECK9-NEXT:    [[TMP290:%.*]] = bitcast i8** [[TMP289]] to i32**
+// CHECK9-NEXT:    store i32* [[TMP266]], i32** [[TMP290]], align 8
+// CHECK9-NEXT:    [[TMP291:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS87]], i64 0, i64 4
+// CHECK9-NEXT:    store i8* null, i8** [[TMP291]], align 8
+// CHECK9-NEXT:    [[TMP292:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS85]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP293:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS86]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP294:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP294]], i32* [[DOTCAPTURE_EXPR_89]], align 4
+// CHECK9-NEXT:    [[TMP295:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_89]], align 4
+// CHECK9-NEXT:    [[SUB91:%.*]] = sub nsw i32 [[TMP295]], 0
+// CHECK9-NEXT:    [[DIV92:%.*]] = sdiv i32 [[SUB91]], 1
+// CHECK9-NEXT:    [[SUB93:%.*]] = sub nsw i32 [[DIV92]], 1
+// CHECK9-NEXT:    store i32 [[SUB93]], i32* [[DOTCAPTURE_EXPR_90]], align 4
+// CHECK9-NEXT:    [[TMP296:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_90]], align 4
+// CHECK9-NEXT:    [[ADD94:%.*]] = add nsw i32 [[TMP296]], 1
+// CHECK9-NEXT:    [[TMP297:%.*]] = zext i32 [[ADD94]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP297]])
+// CHECK9-NEXT:    [[KERNEL_ARGS95:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP298:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP298]], align 4
+// CHECK9-NEXT:    [[TMP299:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP299]], align 4
+// CHECK9-NEXT:    [[TMP300:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP292]], i8*** [[TMP300]], align 8
+// CHECK9-NEXT:    [[TMP301:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP293]], i8*** [[TMP301]], align 8
+// CHECK9-NEXT:    [[TMP302:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.52, i32 0, i32 0), i64** [[TMP302]], align 8
+// CHECK9-NEXT:    [[TMP303:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.53, i32 0, i32 0), i64** [[TMP303]], align 8
+// CHECK9-NEXT:    [[TMP304:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP304]], align 8
+// CHECK9-NEXT:    [[TMP305:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP305]], align 8
+// CHECK9-NEXT:    [[TMP306:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l90.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS95]])
+// CHECK9-NEXT:    [[TMP307:%.*]] = icmp ne i32 [[TMP306]], 0
+// CHECK9-NEXT:    br i1 [[TMP307]], label [[OMP_OFFLOAD_FAILED96:%.*]], label [[OMP_OFFLOAD_CONT97:%.*]]
+// CHECK9:       omp_offload.failed96:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l90(i64 [[TMP261]], i64 [[TMP263]], i32* [[TMP264]], i32* [[TMP265]], i32* [[TMP266]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT97]]
+// CHECK9:       omp_offload.cont97:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -9450,43 +9688,43 @@ int main() {
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_8:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[CH_CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED16:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_21:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED17:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS18:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS19:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS20:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP21:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED29:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS30:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS31:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS32:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP33:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_34:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_35:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[CH_CASTED42:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED43:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS44:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS45:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS46:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP47:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_48:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_49:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED56:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS57:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS58:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS59:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP60:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_61:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_62:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[CH_CASTED69:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED70:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS71:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS72:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS73:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP74:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_75:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_76:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_23:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED31:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS32:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS33:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS34:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP35:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_36:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_37:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[CH_CASTED45:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED46:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS47:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS48:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS49:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP50:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_51:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_52:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED60:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS61:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS62:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS63:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP64:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_65:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_66:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[CH_CASTED74:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED75:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS76:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS77:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS78:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP79:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_80:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_81:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    store i32 10000, i32* [[N]], align 4
 // CHECK11-NEXT:    store i32 100, i32* [[CH]], align 4
@@ -9541,394 +9779,513 @@ int main() {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP29]], 1
 // CHECK11-NEXT:    [[TMP30:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP30]])
-// CHECK11-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l368.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK11-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP25]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP26]], i8*** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l368.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK11-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l368(i32 [[TMP1]], double* [[TMP2]], double* [[TMP3]], double* [[TMP4]]) #[[ATTR2:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP34:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = load double*, double** [[A]], align 4
-// CHECK11-NEXT:    [[TMP36:%.*]] = load double*, double** [[B]], align 4
-// CHECK11-NEXT:    [[TMP37:%.*]] = load double*, double** [[C]], align 4
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP34]], i32* [[TMP39]], align 4
-// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP34]], i32* [[TMP41]], align 4
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP42]], align 4
-// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to double**
-// CHECK11-NEXT:    store double* [[TMP35]], double** [[TMP44]], align 4
-// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to double**
-// CHECK11-NEXT:    store double* [[TMP35]], double** [[TMP46]], align 4
-// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP47]], align 4
-// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to double**
-// CHECK11-NEXT:    store double* [[TMP36]], double** [[TMP49]], align 4
-// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to double**
-// CHECK11-NEXT:    store double* [[TMP36]], double** [[TMP51]], align 4
-// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = load double*, double** [[A]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = load double*, double** [[B]], align 4
+// CHECK11-NEXT:    [[TMP45:%.*]] = load double*, double** [[C]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP42]], i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP42]], i32* [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to double**
+// CHECK11-NEXT:    store double* [[TMP43]], double** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
 // CHECK11-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to double**
-// CHECK11-NEXT:    store double* [[TMP37]], double** [[TMP54]], align 4
-// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to double**
-// CHECK11-NEXT:    store double* [[TMP37]], double** [[TMP56]], align 4
-// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP57]], align 4
-// CHECK11-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP60:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP60]], i32* [[DOTCAPTURE_EXPR_8]], align 4
-// CHECK11-NEXT:    [[TMP61:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_8]], align 4
-// CHECK11-NEXT:    [[SUB10:%.*]] = sub nsw i32 [[TMP61]], 0
+// CHECK11-NEXT:    store double* [[TMP43]], double** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP55]], align 4
+// CHECK11-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to double**
+// CHECK11-NEXT:    store double* [[TMP44]], double** [[TMP57]], align 4
+// CHECK11-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to double**
+// CHECK11-NEXT:    store double* [[TMP44]], double** [[TMP59]], align 4
+// CHECK11-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to double**
+// CHECK11-NEXT:    store double* [[TMP45]], double** [[TMP62]], align 4
+// CHECK11-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP64:%.*]] = bitcast i8** [[TMP63]] to double**
+// CHECK11-NEXT:    store double* [[TMP45]], double** [[TMP64]], align 4
+// CHECK11-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP65]], align 4
+// CHECK11-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP68]], i32* [[DOTCAPTURE_EXPR_8]], align 4
+// CHECK11-NEXT:    [[TMP69:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_8]], align 4
+// CHECK11-NEXT:    [[SUB10:%.*]] = sub nsw i32 [[TMP69]], 0
 // CHECK11-NEXT:    [[DIV11:%.*]] = sdiv i32 [[SUB10]], 1
 // CHECK11-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[DIV11]], 1
 // CHECK11-NEXT:    store i32 [[SUB12]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[TMP62:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[ADD13:%.*]] = add nsw i32 [[TMP62]], 1
-// CHECK11-NEXT:    [[TMP63:%.*]] = zext i32 [[ADD13]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP63]])
-// CHECK11-NEXT:    [[TMP64:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l407.region_id, i32 4, i8** [[TMP58]], i8** [[TMP59]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP65:%.*]] = icmp ne i32 [[TMP64]], 0
-// CHECK11-NEXT:    br i1 [[TMP65]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK11:       omp_offload.failed14:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l407(i32 [[TMP34]], double* [[TMP35]], double* [[TMP36]], double* [[TMP37]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK11:       omp_offload.cont15:
-// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK11-NEXT:    store i32 [[TMP66]], i32* [[CH_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP67:%.*]] = load i32, i32* [[CH_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP68]], i32* [[N_CASTED16]], align 4
-// CHECK11-NEXT:    [[TMP69:%.*]] = load i32, i32* [[N_CASTED16]], align 4
-// CHECK11-NEXT:    [[TMP70:%.*]] = load double*, double** [[A]], align 4
-// CHECK11-NEXT:    [[TMP71:%.*]] = load double*, double** [[B]], align 4
-// CHECK11-NEXT:    [[TMP72:%.*]] = load double*, double** [[C]], align 4
-// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP67]], i32* [[TMP74]], align 4
-// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP67]], i32* [[TMP76]], align 4
-// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP77]], align 4
-// CHECK11-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP69]], i32* [[TMP79]], align 4
-// CHECK11-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP69]], i32* [[TMP81]], align 4
-// CHECK11-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP82]], align 4
-// CHECK11-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to double**
-// CHECK11-NEXT:    store double* [[TMP70]], double** [[TMP84]], align 4
-// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to double**
-// CHECK11-NEXT:    store double* [[TMP70]], double** [[TMP86]], align 4
-// CHECK11-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP87]], align 4
-// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to double**
-// CHECK11-NEXT:    store double* [[TMP71]], double** [[TMP89]], align 4
-// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to double**
-// CHECK11-NEXT:    store double* [[TMP71]], double** [[TMP91]], align 4
-// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP92]], align 4
-// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to double**
-// CHECK11-NEXT:    store double* [[TMP72]], double** [[TMP94]], align 4
-// CHECK11-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to double**
-// CHECK11-NEXT:    store double* [[TMP72]], double** [[TMP96]], align 4
-// CHECK11-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 4
-// CHECK11-NEXT:    store i8* null, i8** [[TMP97]], align 4
-// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP100]], i32* [[DOTCAPTURE_EXPR_21]], align 4
-// CHECK11-NEXT:    [[TMP101:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_21]], align 4
-// CHECK11-NEXT:    [[SUB23:%.*]] = sub nsw i32 [[TMP101]], 0
-// CHECK11-NEXT:    [[DIV24:%.*]] = sdiv i32 [[SUB23]], 1
-// CHECK11-NEXT:    [[SUB25:%.*]] = sub nsw i32 [[DIV24]], 1
-// CHECK11-NEXT:    store i32 [[SUB25]], i32* [[DOTCAPTURE_EXPR_22]], align 4
-// CHECK11-NEXT:    [[TMP102:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
-// CHECK11-NEXT:    [[ADD26:%.*]] = add nsw i32 [[TMP102]], 1
-// CHECK11-NEXT:    [[TMP103:%.*]] = zext i32 [[ADD26]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP103]])
-// CHECK11-NEXT:    [[TMP104:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l446.region_id, i32 5, i8** [[TMP98]], i8** [[TMP99]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP105:%.*]] = icmp ne i32 [[TMP104]], 0
-// CHECK11-NEXT:    br i1 [[TMP105]], label [[OMP_OFFLOAD_FAILED27:%.*]], label [[OMP_OFFLOAD_CONT28:%.*]]
-// CHECK11:       omp_offload.failed27:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l446(i32 [[TMP67]], i32 [[TMP69]], double* [[TMP70]], double* [[TMP71]], double* [[TMP72]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT28]]
-// CHECK11:       omp_offload.cont28:
-// CHECK11-NEXT:    [[TMP106:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP106]], i32* [[N_CASTED29]], align 4
-// CHECK11-NEXT:    [[TMP107:%.*]] = load i32, i32* [[N_CASTED29]], align 4
-// CHECK11-NEXT:    [[TMP108:%.*]] = load double*, double** [[A]], align 4
-// CHECK11-NEXT:    [[TMP109:%.*]] = load double*, double** [[B]], align 4
-// CHECK11-NEXT:    [[TMP110:%.*]] = load double*, double** [[C]], align 4
-// CHECK11-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP107]], i32* [[TMP112]], align 4
-// CHECK11-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP107]], i32* [[TMP114]], align 4
-// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP115]], align 4
-// CHECK11-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to double**
-// CHECK11-NEXT:    store double* [[TMP108]], double** [[TMP117]], align 4
-// CHECK11-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP119:%.*]] = bitcast i8** [[TMP118]] to double**
-// CHECK11-NEXT:    store double* [[TMP108]], double** [[TMP119]], align 4
-// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP120]], align 4
-// CHECK11-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to double**
-// CHECK11-NEXT:    store double* [[TMP109]], double** [[TMP122]], align 4
-// CHECK11-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to double**
-// CHECK11-NEXT:    store double* [[TMP109]], double** [[TMP124]], align 4
-// CHECK11-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP125]], align 4
-// CHECK11-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to double**
-// CHECK11-NEXT:    store double* [[TMP110]], double** [[TMP127]], align 4
-// CHECK11-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to double**
-// CHECK11-NEXT:    store double* [[TMP110]], double** [[TMP129]], align 4
-// CHECK11-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP130]], align 4
-// CHECK11-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP133:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP133]], i32* [[DOTCAPTURE_EXPR_34]], align 4
-// CHECK11-NEXT:    [[TMP134:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_34]], align 4
-// CHECK11-NEXT:    [[SUB36:%.*]] = sub nsw i32 [[TMP134]], 0
-// CHECK11-NEXT:    [[DIV37:%.*]] = sdiv i32 [[SUB36]], 1
-// CHECK11-NEXT:    [[SUB38:%.*]] = sub nsw i32 [[DIV37]], 1
-// CHECK11-NEXT:    store i32 [[SUB38]], i32* [[DOTCAPTURE_EXPR_35]], align 4
-// CHECK11-NEXT:    [[TMP135:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_35]], align 4
-// CHECK11-NEXT:    [[ADD39:%.*]] = add nsw i32 [[TMP135]], 1
-// CHECK11-NEXT:    [[TMP136:%.*]] = zext i32 [[ADD39]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP136]])
-// CHECK11-NEXT:    [[TMP137:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l477.region_id, i32 4, i8** [[TMP131]], i8** [[TMP132]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP138:%.*]] = icmp ne i32 [[TMP137]], 0
-// CHECK11-NEXT:    br i1 [[TMP138]], label [[OMP_OFFLOAD_FAILED40:%.*]], label [[OMP_OFFLOAD_CONT41:%.*]]
-// CHECK11:       omp_offload.failed40:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l477(i32 [[TMP107]], double* [[TMP108]], double* [[TMP109]], double* [[TMP110]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT41]]
-// CHECK11:       omp_offload.cont41:
-// CHECK11-NEXT:    [[TMP139:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK11-NEXT:    store i32 [[TMP139]], i32* [[CH_CASTED42]], align 4
-// CHECK11-NEXT:    [[TMP140:%.*]] = load i32, i32* [[CH_CASTED42]], align 4
-// CHECK11-NEXT:    [[TMP141:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP141]], i32* [[N_CASTED43]], align 4
-// CHECK11-NEXT:    [[TMP142:%.*]] = load i32, i32* [[N_CASTED43]], align 4
-// CHECK11-NEXT:    [[TMP143:%.*]] = load double*, double** [[A]], align 4
-// CHECK11-NEXT:    [[TMP144:%.*]] = load double*, double** [[B]], align 4
-// CHECK11-NEXT:    [[TMP145:%.*]] = load double*, double** [[C]], align 4
-// CHECK11-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP140]], i32* [[TMP147]], align 4
-// CHECK11-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP140]], i32* [[TMP149]], align 4
-// CHECK11-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP150]], align 4
-// CHECK11-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP142]], i32* [[TMP152]], align 4
-// CHECK11-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP142]], i32* [[TMP154]], align 4
-// CHECK11-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP155]], align 4
-// CHECK11-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to double**
-// CHECK11-NEXT:    store double* [[TMP143]], double** [[TMP157]], align 4
-// CHECK11-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to double**
-// CHECK11-NEXT:    store double* [[TMP143]], double** [[TMP159]], align 4
-// CHECK11-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP160]], align 4
-// CHECK11-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP162:%.*]] = bitcast i8** [[TMP161]] to double**
-// CHECK11-NEXT:    store double* [[TMP144]], double** [[TMP162]], align 4
-// CHECK11-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP164:%.*]] = bitcast i8** [[TMP163]] to double**
-// CHECK11-NEXT:    store double* [[TMP144]], double** [[TMP164]], align 4
-// CHECK11-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP165]], align 4
-// CHECK11-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to double**
-// CHECK11-NEXT:    store double* [[TMP145]], double** [[TMP167]], align 4
-// CHECK11-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to double**
-// CHECK11-NEXT:    store double* [[TMP145]], double** [[TMP169]], align 4
-// CHECK11-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 4
-// CHECK11-NEXT:    store i8* null, i8** [[TMP170]], align 4
-// CHECK11-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP70:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[ADD13:%.*]] = add nsw i32 [[TMP70]], 1
+// CHECK11-NEXT:    [[TMP71:%.*]] = zext i32 [[ADD13]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP71]])
+// CHECK11-NEXT:    [[KERNEL_ARGS14:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP72]], align 4
+// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP73]], align 4
+// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP66]], i8*** [[TMP74]], align 4
+// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP67]], i8*** [[TMP75]], align 4
+// CHECK11-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP76]], align 4
+// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP77]], align 4
+// CHECK11-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP78]], align 4
+// CHECK11-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP79]], align 4
+// CHECK11-NEXT:    [[TMP80:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l407.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]])
+// CHECK11-NEXT:    [[TMP81:%.*]] = icmp ne i32 [[TMP80]], 0
+// CHECK11-NEXT:    br i1 [[TMP81]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
+// CHECK11:       omp_offload.failed15:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l407(i32 [[TMP42]], double* [[TMP43]], double* [[TMP44]], double* [[TMP45]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
+// CHECK11:       omp_offload.cont16:
+// CHECK11-NEXT:    [[TMP82:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK11-NEXT:    store i32 [[TMP82]], i32* [[CH_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP83:%.*]] = load i32, i32* [[CH_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP84:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP84]], i32* [[N_CASTED17]], align 4
+// CHECK11-NEXT:    [[TMP85:%.*]] = load i32, i32* [[N_CASTED17]], align 4
+// CHECK11-NEXT:    [[TMP86:%.*]] = load double*, double** [[A]], align 4
+// CHECK11-NEXT:    [[TMP87:%.*]] = load double*, double** [[B]], align 4
+// CHECK11-NEXT:    [[TMP88:%.*]] = load double*, double** [[C]], align 4
+// CHECK11-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP83]], i32* [[TMP90]], align 4
+// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP83]], i32* [[TMP92]], align 4
+// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP93]], align 4
+// CHECK11-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP85]], i32* [[TMP95]], align 4
+// CHECK11-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP85]], i32* [[TMP97]], align 4
+// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP98]], align 4
+// CHECK11-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP100:%.*]] = bitcast i8** [[TMP99]] to double**
+// CHECK11-NEXT:    store double* [[TMP86]], double** [[TMP100]], align 4
+// CHECK11-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to double**
+// CHECK11-NEXT:    store double* [[TMP86]], double** [[TMP102]], align 4
+// CHECK11-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP103]], align 4
+// CHECK11-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to double**
+// CHECK11-NEXT:    store double* [[TMP87]], double** [[TMP105]], align 4
+// CHECK11-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to double**
+// CHECK11-NEXT:    store double* [[TMP87]], double** [[TMP107]], align 4
+// CHECK11-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP108]], align 4
+// CHECK11-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to double**
+// CHECK11-NEXT:    store double* [[TMP88]], double** [[TMP110]], align 4
+// CHECK11-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to double**
+// CHECK11-NEXT:    store double* [[TMP88]], double** [[TMP112]], align 4
+// CHECK11-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 4
+// CHECK11-NEXT:    store i8* null, i8** [[TMP113]], align 4
+// CHECK11-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP116:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP116]], i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK11-NEXT:    [[TMP117:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK11-NEXT:    [[SUB24:%.*]] = sub nsw i32 [[TMP117]], 0
+// CHECK11-NEXT:    [[DIV25:%.*]] = sdiv i32 [[SUB24]], 1
+// CHECK11-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[DIV25]], 1
+// CHECK11-NEXT:    store i32 [[SUB26]], i32* [[DOTCAPTURE_EXPR_23]], align 4
+// CHECK11-NEXT:    [[TMP118:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_23]], align 4
+// CHECK11-NEXT:    [[ADD27:%.*]] = add nsw i32 [[TMP118]], 1
+// CHECK11-NEXT:    [[TMP119:%.*]] = zext i32 [[ADD27]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP119]])
+// CHECK11-NEXT:    [[KERNEL_ARGS28:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP120]], align 4
+// CHECK11-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP121]], align 4
+// CHECK11-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP114]], i8*** [[TMP122]], align 4
+// CHECK11-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP115]], i8*** [[TMP123]], align 4
+// CHECK11-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP124]], align 4
+// CHECK11-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP125]], align 4
+// CHECK11-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP126]], align 4
+// CHECK11-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP127]], align 4
+// CHECK11-NEXT:    [[TMP128:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l446.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]])
+// CHECK11-NEXT:    [[TMP129:%.*]] = icmp ne i32 [[TMP128]], 0
+// CHECK11-NEXT:    br i1 [[TMP129]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
+// CHECK11:       omp_offload.failed29:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l446(i32 [[TMP83]], i32 [[TMP85]], double* [[TMP86]], double* [[TMP87]], double* [[TMP88]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
+// CHECK11:       omp_offload.cont30:
+// CHECK11-NEXT:    [[TMP130:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP130]], i32* [[N_CASTED31]], align 4
+// CHECK11-NEXT:    [[TMP131:%.*]] = load i32, i32* [[N_CASTED31]], align 4
+// CHECK11-NEXT:    [[TMP132:%.*]] = load double*, double** [[A]], align 4
+// CHECK11-NEXT:    [[TMP133:%.*]] = load double*, double** [[B]], align 4
+// CHECK11-NEXT:    [[TMP134:%.*]] = load double*, double** [[C]], align 4
+// CHECK11-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP131]], i32* [[TMP136]], align 4
+// CHECK11-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP131]], i32* [[TMP138]], align 4
+// CHECK11-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP139]], align 4
+// CHECK11-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to double**
+// CHECK11-NEXT:    store double* [[TMP132]], double** [[TMP141]], align 4
+// CHECK11-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to double**
+// CHECK11-NEXT:    store double* [[TMP132]], double** [[TMP143]], align 4
+// CHECK11-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP144]], align 4
+// CHECK11-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to double**
+// CHECK11-NEXT:    store double* [[TMP133]], double** [[TMP146]], align 4
+// CHECK11-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to double**
+// CHECK11-NEXT:    store double* [[TMP133]], double** [[TMP148]], align 4
+// CHECK11-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP149]], align 4
+// CHECK11-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to double**
+// CHECK11-NEXT:    store double* [[TMP134]], double** [[TMP151]], align 4
+// CHECK11-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to double**
+// CHECK11-NEXT:    store double* [[TMP134]], double** [[TMP153]], align 4
+// CHECK11-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP154]], align 4
+// CHECK11-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP157:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP157]], i32* [[DOTCAPTURE_EXPR_36]], align 4
+// CHECK11-NEXT:    [[TMP158:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_36]], align 4
+// CHECK11-NEXT:    [[SUB38:%.*]] = sub nsw i32 [[TMP158]], 0
+// CHECK11-NEXT:    [[DIV39:%.*]] = sdiv i32 [[SUB38]], 1
+// CHECK11-NEXT:    [[SUB40:%.*]] = sub nsw i32 [[DIV39]], 1
+// CHECK11-NEXT:    store i32 [[SUB40]], i32* [[DOTCAPTURE_EXPR_37]], align 4
+// CHECK11-NEXT:    [[TMP159:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_37]], align 4
+// CHECK11-NEXT:    [[ADD41:%.*]] = add nsw i32 [[TMP159]], 1
+// CHECK11-NEXT:    [[TMP160:%.*]] = zext i32 [[ADD41]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP160]])
+// CHECK11-NEXT:    [[KERNEL_ARGS42:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP161]], align 4
+// CHECK11-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP162]], align 4
+// CHECK11-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP155]], i8*** [[TMP163]], align 4
+// CHECK11-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP156]], i8*** [[TMP164]], align 4
+// CHECK11-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP165]], align 4
+// CHECK11-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP166]], align 4
+// CHECK11-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP167]], align 4
+// CHECK11-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP168]], align 4
+// CHECK11-NEXT:    [[TMP169:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l477.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]])
+// CHECK11-NEXT:    [[TMP170:%.*]] = icmp ne i32 [[TMP169]], 0
+// CHECK11-NEXT:    br i1 [[TMP170]], label [[OMP_OFFLOAD_FAILED43:%.*]], label [[OMP_OFFLOAD_CONT44:%.*]]
+// CHECK11:       omp_offload.failed43:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l477(i32 [[TMP131]], double* [[TMP132]], double* [[TMP133]], double* [[TMP134]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT44]]
+// CHECK11:       omp_offload.cont44:
+// CHECK11-NEXT:    [[TMP171:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK11-NEXT:    store i32 [[TMP171]], i32* [[CH_CASTED45]], align 4
+// CHECK11-NEXT:    [[TMP172:%.*]] = load i32, i32* [[CH_CASTED45]], align 4
 // CHECK11-NEXT:    [[TMP173:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP173]], i32* [[DOTCAPTURE_EXPR_48]], align 4
-// CHECK11-NEXT:    [[TMP174:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_48]], align 4
-// CHECK11-NEXT:    [[SUB50:%.*]] = sub nsw i32 [[TMP174]], 0
-// CHECK11-NEXT:    [[DIV51:%.*]] = sdiv i32 [[SUB50]], 1
-// CHECK11-NEXT:    [[SUB52:%.*]] = sub nsw i32 [[DIV51]], 1
-// CHECK11-NEXT:    store i32 [[SUB52]], i32* [[DOTCAPTURE_EXPR_49]], align 4
-// CHECK11-NEXT:    [[TMP175:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_49]], align 4
-// CHECK11-NEXT:    [[ADD53:%.*]] = add nsw i32 [[TMP175]], 1
-// CHECK11-NEXT:    [[TMP176:%.*]] = zext i32 [[ADD53]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP176]])
-// CHECK11-NEXT:    [[TMP177:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l505.region_id, i32 5, i8** [[TMP171]], i8** [[TMP172]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP178:%.*]] = icmp ne i32 [[TMP177]], 0
-// CHECK11-NEXT:    br i1 [[TMP178]], label [[OMP_OFFLOAD_FAILED54:%.*]], label [[OMP_OFFLOAD_CONT55:%.*]]
-// CHECK11:       omp_offload.failed54:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l505(i32 [[TMP140]], i32 [[TMP142]], double* [[TMP143]], double* [[TMP144]], double* [[TMP145]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT55]]
-// CHECK11:       omp_offload.cont55:
-// CHECK11-NEXT:    [[TMP179:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP179]], i32* [[N_CASTED56]], align 4
-// CHECK11-NEXT:    [[TMP180:%.*]] = load i32, i32* [[N_CASTED56]], align 4
-// CHECK11-NEXT:    [[TMP181:%.*]] = load double*, double** [[A]], align 4
-// CHECK11-NEXT:    [[TMP182:%.*]] = load double*, double** [[B]], align 4
-// CHECK11-NEXT:    [[TMP183:%.*]] = load double*, double** [[C]], align 4
-// CHECK11-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP185:%.*]] = bitcast i8** [[TMP184]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP180]], i32* [[TMP185]], align 4
-// CHECK11-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP187:%.*]] = bitcast i8** [[TMP186]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP180]], i32* [[TMP187]], align 4
-// CHECK11-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP188]], align 4
-// CHECK11-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP190:%.*]] = bitcast i8** [[TMP189]] to double**
-// CHECK11-NEXT:    store double* [[TMP181]], double** [[TMP190]], align 4
-// CHECK11-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP192:%.*]] = bitcast i8** [[TMP191]] to double**
-// CHECK11-NEXT:    store double* [[TMP181]], double** [[TMP192]], align 4
-// CHECK11-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP193]], align 4
-// CHECK11-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP195:%.*]] = bitcast i8** [[TMP194]] to double**
-// CHECK11-NEXT:    store double* [[TMP182]], double** [[TMP195]], align 4
-// CHECK11-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP197:%.*]] = bitcast i8** [[TMP196]] to double**
-// CHECK11-NEXT:    store double* [[TMP182]], double** [[TMP197]], align 4
-// CHECK11-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP198]], align 4
-// CHECK11-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP200:%.*]] = bitcast i8** [[TMP199]] to double**
-// CHECK11-NEXT:    store double* [[TMP183]], double** [[TMP200]], align 4
-// CHECK11-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP202:%.*]] = bitcast i8** [[TMP201]] to double**
-// CHECK11-NEXT:    store double* [[TMP183]], double** [[TMP202]], align 4
-// CHECK11-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP203]], align 4
-// CHECK11-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP206:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP206]], i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK11-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK11-NEXT:    [[SUB63:%.*]] = sub nsw i32 [[TMP207]], 0
-// CHECK11-NEXT:    [[DIV64:%.*]] = sdiv i32 [[SUB63]], 1
-// CHECK11-NEXT:    [[SUB65:%.*]] = sub nsw i32 [[DIV64]], 1
-// CHECK11-NEXT:    store i32 [[SUB65]], i32* [[DOTCAPTURE_EXPR_62]], align 4
-// CHECK11-NEXT:    [[TMP208:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_62]], align 4
-// CHECK11-NEXT:    [[ADD66:%.*]] = add nsw i32 [[TMP208]], 1
-// CHECK11-NEXT:    [[TMP209:%.*]] = zext i32 [[ADD66]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP209]])
-// CHECK11-NEXT:    [[TMP210:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l535.region_id, i32 4, i8** [[TMP204]], i8** [[TMP205]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP211:%.*]] = icmp ne i32 [[TMP210]], 0
-// CHECK11-NEXT:    br i1 [[TMP211]], label [[OMP_OFFLOAD_FAILED67:%.*]], label [[OMP_OFFLOAD_CONT68:%.*]]
-// CHECK11:       omp_offload.failed67:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l535(i32 [[TMP180]], double* [[TMP181]], double* [[TMP182]], double* [[TMP183]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT68]]
-// CHECK11:       omp_offload.cont68:
-// CHECK11-NEXT:    [[TMP212:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK11-NEXT:    store i32 [[TMP212]], i32* [[CH_CASTED69]], align 4
-// CHECK11-NEXT:    [[TMP213:%.*]] = load i32, i32* [[CH_CASTED69]], align 4
-// CHECK11-NEXT:    [[TMP214:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP214]], i32* [[N_CASTED70]], align 4
-// CHECK11-NEXT:    [[TMP215:%.*]] = load i32, i32* [[N_CASTED70]], align 4
-// CHECK11-NEXT:    [[TMP216:%.*]] = load double*, double** [[A]], align 4
-// CHECK11-NEXT:    [[TMP217:%.*]] = load double*, double** [[B]], align 4
-// CHECK11-NEXT:    [[TMP218:%.*]] = load double*, double** [[C]], align 4
-// CHECK11-NEXT:    [[TMP219:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP220:%.*]] = bitcast i8** [[TMP219]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP213]], i32* [[TMP220]], align 4
-// CHECK11-NEXT:    [[TMP221:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP222:%.*]] = bitcast i8** [[TMP221]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP213]], i32* [[TMP222]], align 4
-// CHECK11-NEXT:    [[TMP223:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP223]], align 4
-// CHECK11-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 [[TMP173]], i32* [[N_CASTED46]], align 4
+// CHECK11-NEXT:    [[TMP174:%.*]] = load i32, i32* [[N_CASTED46]], align 4
+// CHECK11-NEXT:    [[TMP175:%.*]] = load double*, double** [[A]], align 4
+// CHECK11-NEXT:    [[TMP176:%.*]] = load double*, double** [[B]], align 4
+// CHECK11-NEXT:    [[TMP177:%.*]] = load double*, double** [[C]], align 4
+// CHECK11-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP172]], i32* [[TMP179]], align 4
+// CHECK11-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP172]], i32* [[TMP181]], align 4
+// CHECK11-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP182]], align 4
+// CHECK11-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP184:%.*]] = bitcast i8** [[TMP183]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP174]], i32* [[TMP184]], align 4
+// CHECK11-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP186:%.*]] = bitcast i8** [[TMP185]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP174]], i32* [[TMP186]], align 4
+// CHECK11-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP187]], align 4
+// CHECK11-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP189:%.*]] = bitcast i8** [[TMP188]] to double**
+// CHECK11-NEXT:    store double* [[TMP175]], double** [[TMP189]], align 4
+// CHECK11-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP191:%.*]] = bitcast i8** [[TMP190]] to double**
+// CHECK11-NEXT:    store double* [[TMP175]], double** [[TMP191]], align 4
+// CHECK11-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP192]], align 4
+// CHECK11-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP194:%.*]] = bitcast i8** [[TMP193]] to double**
+// CHECK11-NEXT:    store double* [[TMP176]], double** [[TMP194]], align 4
+// CHECK11-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP196:%.*]] = bitcast i8** [[TMP195]] to double**
+// CHECK11-NEXT:    store double* [[TMP176]], double** [[TMP196]], align 4
+// CHECK11-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP197]], align 4
+// CHECK11-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP199:%.*]] = bitcast i8** [[TMP198]] to double**
+// CHECK11-NEXT:    store double* [[TMP177]], double** [[TMP199]], align 4
+// CHECK11-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP201:%.*]] = bitcast i8** [[TMP200]] to double**
+// CHECK11-NEXT:    store double* [[TMP177]], double** [[TMP201]], align 4
+// CHECK11-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 4
+// CHECK11-NEXT:    store i8* null, i8** [[TMP202]], align 4
+// CHECK11-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP205:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP205]], i32* [[DOTCAPTURE_EXPR_51]], align 4
+// CHECK11-NEXT:    [[TMP206:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_51]], align 4
+// CHECK11-NEXT:    [[SUB53:%.*]] = sub nsw i32 [[TMP206]], 0
+// CHECK11-NEXT:    [[DIV54:%.*]] = sdiv i32 [[SUB53]], 1
+// CHECK11-NEXT:    [[SUB55:%.*]] = sub nsw i32 [[DIV54]], 1
+// CHECK11-NEXT:    store i32 [[SUB55]], i32* [[DOTCAPTURE_EXPR_52]], align 4
+// CHECK11-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_52]], align 4
+// CHECK11-NEXT:    [[ADD56:%.*]] = add nsw i32 [[TMP207]], 1
+// CHECK11-NEXT:    [[TMP208:%.*]] = zext i32 [[ADD56]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP208]])
+// CHECK11-NEXT:    [[KERNEL_ARGS57:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP209:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP209]], align 4
+// CHECK11-NEXT:    [[TMP210:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP210]], align 4
+// CHECK11-NEXT:    [[TMP211:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP203]], i8*** [[TMP211]], align 4
+// CHECK11-NEXT:    [[TMP212:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP204]], i8*** [[TMP212]], align 4
+// CHECK11-NEXT:    [[TMP213:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP213]], align 4
+// CHECK11-NEXT:    [[TMP214:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP214]], align 4
+// CHECK11-NEXT:    [[TMP215:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP215]], align 4
+// CHECK11-NEXT:    [[TMP216:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP216]], align 4
+// CHECK11-NEXT:    [[TMP217:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l505.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]])
+// CHECK11-NEXT:    [[TMP218:%.*]] = icmp ne i32 [[TMP217]], 0
+// CHECK11-NEXT:    br i1 [[TMP218]], label [[OMP_OFFLOAD_FAILED58:%.*]], label [[OMP_OFFLOAD_CONT59:%.*]]
+// CHECK11:       omp_offload.failed58:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l505(i32 [[TMP172]], i32 [[TMP174]], double* [[TMP175]], double* [[TMP176]], double* [[TMP177]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT59]]
+// CHECK11:       omp_offload.cont59:
+// CHECK11-NEXT:    [[TMP219:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP219]], i32* [[N_CASTED60]], align 4
+// CHECK11-NEXT:    [[TMP220:%.*]] = load i32, i32* [[N_CASTED60]], align 4
+// CHECK11-NEXT:    [[TMP221:%.*]] = load double*, double** [[A]], align 4
+// CHECK11-NEXT:    [[TMP222:%.*]] = load double*, double** [[B]], align 4
+// CHECK11-NEXT:    [[TMP223:%.*]] = load double*, double** [[C]], align 4
+// CHECK11-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP225:%.*]] = bitcast i8** [[TMP224]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP215]], i32* [[TMP225]], align 4
-// CHECK11-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 [[TMP220]], i32* [[TMP225]], align 4
+// CHECK11-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP227:%.*]] = bitcast i8** [[TMP226]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP215]], i32* [[TMP227]], align 4
-// CHECK11-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 [[TMP220]], i32* [[TMP227]], align 4
+// CHECK11-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 0
 // CHECK11-NEXT:    store i8* null, i8** [[TMP228]], align 4
-// CHECK11-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 1
 // CHECK11-NEXT:    [[TMP230:%.*]] = bitcast i8** [[TMP229]] to double**
-// CHECK11-NEXT:    store double* [[TMP216]], double** [[TMP230]], align 4
-// CHECK11-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 2
+// CHECK11-NEXT:    store double* [[TMP221]], double** [[TMP230]], align 4
+// CHECK11-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 1
 // CHECK11-NEXT:    [[TMP232:%.*]] = bitcast i8** [[TMP231]] to double**
-// CHECK11-NEXT:    store double* [[TMP216]], double** [[TMP232]], align 4
-// CHECK11-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 2
+// CHECK11-NEXT:    store double* [[TMP221]], double** [[TMP232]], align 4
+// CHECK11-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 1
 // CHECK11-NEXT:    store i8* null, i8** [[TMP233]], align 4
-// CHECK11-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 2
 // CHECK11-NEXT:    [[TMP235:%.*]] = bitcast i8** [[TMP234]] to double**
-// CHECK11-NEXT:    store double* [[TMP217]], double** [[TMP235]], align 4
-// CHECK11-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 3
+// CHECK11-NEXT:    store double* [[TMP222]], double** [[TMP235]], align 4
+// CHECK11-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 2
 // CHECK11-NEXT:    [[TMP237:%.*]] = bitcast i8** [[TMP236]] to double**
-// CHECK11-NEXT:    store double* [[TMP217]], double** [[TMP237]], align 4
-// CHECK11-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 3
+// CHECK11-NEXT:    store double* [[TMP222]], double** [[TMP237]], align 4
+// CHECK11-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 2
 // CHECK11-NEXT:    store i8* null, i8** [[TMP238]], align 4
-// CHECK11-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 3
 // CHECK11-NEXT:    [[TMP240:%.*]] = bitcast i8** [[TMP239]] to double**
-// CHECK11-NEXT:    store double* [[TMP218]], double** [[TMP240]], align 4
-// CHECK11-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 4
+// CHECK11-NEXT:    store double* [[TMP223]], double** [[TMP240]], align 4
+// CHECK11-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 3
 // CHECK11-NEXT:    [[TMP242:%.*]] = bitcast i8** [[TMP241]] to double**
-// CHECK11-NEXT:    store double* [[TMP218]], double** [[TMP242]], align 4
-// CHECK11-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 4
+// CHECK11-NEXT:    store double* [[TMP223]], double** [[TMP242]], align 4
+// CHECK11-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 3
 // CHECK11-NEXT:    store i8* null, i8** [[TMP243]], align 4
-// CHECK11-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP246:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_75]], align 4
-// CHECK11-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_75]], align 4
-// CHECK11-NEXT:    [[SUB77:%.*]] = sub nsw i32 [[TMP247]], 0
-// CHECK11-NEXT:    [[DIV78:%.*]] = sdiv i32 [[SUB77]], 1
-// CHECK11-NEXT:    [[SUB79:%.*]] = sub nsw i32 [[DIV78]], 1
-// CHECK11-NEXT:    store i32 [[SUB79]], i32* [[DOTCAPTURE_EXPR_76]], align 4
-// CHECK11-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_76]], align 4
-// CHECK11-NEXT:    [[ADD80:%.*]] = add nsw i32 [[TMP248]], 1
-// CHECK11-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD80]] to i64
+// CHECK11-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_65]], align 4
+// CHECK11-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_65]], align 4
+// CHECK11-NEXT:    [[SUB67:%.*]] = sub nsw i32 [[TMP247]], 0
+// CHECK11-NEXT:    [[DIV68:%.*]] = sdiv i32 [[SUB67]], 1
+// CHECK11-NEXT:    [[SUB69:%.*]] = sub nsw i32 [[DIV68]], 1
+// CHECK11-NEXT:    store i32 [[SUB69]], i32* [[DOTCAPTURE_EXPR_66]], align 4
+// CHECK11-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_66]], align 4
+// CHECK11-NEXT:    [[ADD70:%.*]] = add nsw i32 [[TMP248]], 1
+// CHECK11-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD70]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP249]])
-// CHECK11-NEXT:    [[TMP250:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l561.region_id, i32 5, i8** [[TMP244]], i8** [[TMP245]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP251:%.*]] = icmp ne i32 [[TMP250]], 0
-// CHECK11-NEXT:    br i1 [[TMP251]], label [[OMP_OFFLOAD_FAILED81:%.*]], label [[OMP_OFFLOAD_CONT82:%.*]]
-// CHECK11:       omp_offload.failed81:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l561(i32 [[TMP213]], i32 [[TMP215]], double* [[TMP216]], double* [[TMP217]], double* [[TMP218]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT82]]
-// CHECK11:       omp_offload.cont82:
+// CHECK11-NEXT:    [[KERNEL_ARGS71:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP250:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP250]], align 4
+// CHECK11-NEXT:    [[TMP251:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP251]], align 4
+// CHECK11-NEXT:    [[TMP252:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP244]], i8*** [[TMP252]], align 4
+// CHECK11-NEXT:    [[TMP253:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP245]], i8*** [[TMP253]], align 4
+// CHECK11-NEXT:    [[TMP254:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP254]], align 4
+// CHECK11-NEXT:    [[TMP255:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP255]], align 4
+// CHECK11-NEXT:    [[TMP256:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP256]], align 4
+// CHECK11-NEXT:    [[TMP257:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP257]], align 4
+// CHECK11-NEXT:    [[TMP258:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l535.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]])
+// CHECK11-NEXT:    [[TMP259:%.*]] = icmp ne i32 [[TMP258]], 0
+// CHECK11-NEXT:    br i1 [[TMP259]], label [[OMP_OFFLOAD_FAILED72:%.*]], label [[OMP_OFFLOAD_CONT73:%.*]]
+// CHECK11:       omp_offload.failed72:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l535(i32 [[TMP220]], double* [[TMP221]], double* [[TMP222]], double* [[TMP223]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT73]]
+// CHECK11:       omp_offload.cont73:
+// CHECK11-NEXT:    [[TMP260:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK11-NEXT:    store i32 [[TMP260]], i32* [[CH_CASTED74]], align 4
+// CHECK11-NEXT:    [[TMP261:%.*]] = load i32, i32* [[CH_CASTED74]], align 4
+// CHECK11-NEXT:    [[TMP262:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP262]], i32* [[N_CASTED75]], align 4
+// CHECK11-NEXT:    [[TMP263:%.*]] = load i32, i32* [[N_CASTED75]], align 4
+// CHECK11-NEXT:    [[TMP264:%.*]] = load double*, double** [[A]], align 4
+// CHECK11-NEXT:    [[TMP265:%.*]] = load double*, double** [[B]], align 4
+// CHECK11-NEXT:    [[TMP266:%.*]] = load double*, double** [[C]], align 4
+// CHECK11-NEXT:    [[TMP267:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP268:%.*]] = bitcast i8** [[TMP267]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP261]], i32* [[TMP268]], align 4
+// CHECK11-NEXT:    [[TMP269:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP270:%.*]] = bitcast i8** [[TMP269]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP261]], i32* [[TMP270]], align 4
+// CHECK11-NEXT:    [[TMP271:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP271]], align 4
+// CHECK11-NEXT:    [[TMP272:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP273:%.*]] = bitcast i8** [[TMP272]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP263]], i32* [[TMP273]], align 4
+// CHECK11-NEXT:    [[TMP274:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP275:%.*]] = bitcast i8** [[TMP274]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP263]], i32* [[TMP275]], align 4
+// CHECK11-NEXT:    [[TMP276:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP276]], align 4
+// CHECK11-NEXT:    [[TMP277:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP278:%.*]] = bitcast i8** [[TMP277]] to double**
+// CHECK11-NEXT:    store double* [[TMP264]], double** [[TMP278]], align 4
+// CHECK11-NEXT:    [[TMP279:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP280:%.*]] = bitcast i8** [[TMP279]] to double**
+// CHECK11-NEXT:    store double* [[TMP264]], double** [[TMP280]], align 4
+// CHECK11-NEXT:    [[TMP281:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP281]], align 4
+// CHECK11-NEXT:    [[TMP282:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP283:%.*]] = bitcast i8** [[TMP282]] to double**
+// CHECK11-NEXT:    store double* [[TMP265]], double** [[TMP283]], align 4
+// CHECK11-NEXT:    [[TMP284:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP285:%.*]] = bitcast i8** [[TMP284]] to double**
+// CHECK11-NEXT:    store double* [[TMP265]], double** [[TMP285]], align 4
+// CHECK11-NEXT:    [[TMP286:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP286]], align 4
+// CHECK11-NEXT:    [[TMP287:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP288:%.*]] = bitcast i8** [[TMP287]] to double**
+// CHECK11-NEXT:    store double* [[TMP266]], double** [[TMP288]], align 4
+// CHECK11-NEXT:    [[TMP289:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP290:%.*]] = bitcast i8** [[TMP289]] to double**
+// CHECK11-NEXT:    store double* [[TMP266]], double** [[TMP290]], align 4
+// CHECK11-NEXT:    [[TMP291:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 4
+// CHECK11-NEXT:    store i8* null, i8** [[TMP291]], align 4
+// CHECK11-NEXT:    [[TMP292:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP293:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP294:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP294]], i32* [[DOTCAPTURE_EXPR_80]], align 4
+// CHECK11-NEXT:    [[TMP295:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_80]], align 4
+// CHECK11-NEXT:    [[SUB82:%.*]] = sub nsw i32 [[TMP295]], 0
+// CHECK11-NEXT:    [[DIV83:%.*]] = sdiv i32 [[SUB82]], 1
+// CHECK11-NEXT:    [[SUB84:%.*]] = sub nsw i32 [[DIV83]], 1
+// CHECK11-NEXT:    store i32 [[SUB84]], i32* [[DOTCAPTURE_EXPR_81]], align 4
+// CHECK11-NEXT:    [[TMP296:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_81]], align 4
+// CHECK11-NEXT:    [[ADD85:%.*]] = add nsw i32 [[TMP296]], 1
+// CHECK11-NEXT:    [[TMP297:%.*]] = zext i32 [[ADD85]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP297]])
+// CHECK11-NEXT:    [[KERNEL_ARGS86:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP298:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP298]], align 4
+// CHECK11-NEXT:    [[TMP299:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP299]], align 4
+// CHECK11-NEXT:    [[TMP300:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP292]], i8*** [[TMP300]], align 4
+// CHECK11-NEXT:    [[TMP301:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP293]], i8*** [[TMP301]], align 4
+// CHECK11-NEXT:    [[TMP302:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP302]], align 4
+// CHECK11-NEXT:    [[TMP303:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP303]], align 4
+// CHECK11-NEXT:    [[TMP304:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP304]], align 4
+// CHECK11-NEXT:    [[TMP305:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP305]], align 4
+// CHECK11-NEXT:    [[TMP306:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l561.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]])
+// CHECK11-NEXT:    [[TMP307:%.*]] = icmp ne i32 [[TMP306]], 0
+// CHECK11-NEXT:    br i1 [[TMP307]], label [[OMP_OFFLOAD_FAILED87:%.*]], label [[OMP_OFFLOAD_CONT88:%.*]]
+// CHECK11:       omp_offload.failed87:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l561(i32 [[TMP261]], i32 [[TMP263]], double* [[TMP264]], double* [[TMP265]], double* [[TMP266]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT88]]
+// CHECK11:       omp_offload.cont88:
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    ret i32 [[CALL]]
 //
@@ -11747,43 +12104,43 @@ int main() {
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_8:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[CH_CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED16:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_21:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED17:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS18:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS19:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS20:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP21:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED29:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS30:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS31:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS32:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP33:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_34:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_35:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[CH_CASTED42:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED43:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS44:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS45:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS46:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP47:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_48:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_49:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED56:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS57:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS58:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS59:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP60:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_61:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_62:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[CH_CASTED69:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED70:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS71:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS72:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS73:%.*]] = alloca [5 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP74:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_75:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_76:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_23:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED31:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS32:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS33:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS34:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP35:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_36:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_37:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[CH_CASTED45:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED46:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS47:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS48:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS49:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP50:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_51:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_52:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED60:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS61:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS62:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS63:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP64:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_65:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_66:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[CH_CASTED74:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED75:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS76:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS77:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS78:%.*]] = alloca [5 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP79:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_80:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_81:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 10000, i32* [[N]], align 4
 // CHECK11-NEXT:    store i32 100, i32* [[CH]], align 4
 // CHECK11-NEXT:    [[TMP0:%.*]] = load i32, i32* [[N]], align 4
@@ -11837,394 +12194,513 @@ int main() {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP29]], 1
 // CHECK11-NEXT:    [[TMP30:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP30]])
-// CHECK11-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l42.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK11-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP25]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP26]], i8*** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l42.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK11-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l42(i32 [[TMP1]], i32* [[TMP2]], i32* [[TMP3]], i32* [[TMP4]]) #[[ATTR2]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP34:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = load i32*, i32** [[A]], align 4
-// CHECK11-NEXT:    [[TMP36:%.*]] = load i32*, i32** [[B]], align 4
-// CHECK11-NEXT:    [[TMP37:%.*]] = load i32*, i32** [[C]], align 4
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP34]], i32* [[TMP39]], align 4
-// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP34]], i32* [[TMP41]], align 4
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP42]], align 4
-// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP35]], i32** [[TMP44]], align 4
-// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP35]], i32** [[TMP46]], align 4
-// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP47]], align 4
-// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP36]], i32** [[TMP49]], align 4
-// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP36]], i32** [[TMP51]], align 4
-// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = load i32*, i32** [[A]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = load i32*, i32** [[B]], align 4
+// CHECK11-NEXT:    [[TMP45:%.*]] = load i32*, i32** [[C]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP42]], i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP42]], i32* [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP43]], i32** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
 // CHECK11-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP37]], i32** [[TMP54]], align 4
-// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP37]], i32** [[TMP56]], align 4
-// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP57]], align 4
-// CHECK11-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP60:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP60]], i32* [[DOTCAPTURE_EXPR_8]], align 4
-// CHECK11-NEXT:    [[TMP61:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_8]], align 4
-// CHECK11-NEXT:    [[SUB10:%.*]] = sub nsw i32 [[TMP61]], 0
+// CHECK11-NEXT:    store i32* [[TMP43]], i32** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP55]], align 4
+// CHECK11-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP44]], i32** [[TMP57]], align 4
+// CHECK11-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP44]], i32** [[TMP59]], align 4
+// CHECK11-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP45]], i32** [[TMP62]], align 4
+// CHECK11-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP64:%.*]] = bitcast i8** [[TMP63]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP45]], i32** [[TMP64]], align 4
+// CHECK11-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP65]], align 4
+// CHECK11-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP68]], i32* [[DOTCAPTURE_EXPR_8]], align 4
+// CHECK11-NEXT:    [[TMP69:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_8]], align 4
+// CHECK11-NEXT:    [[SUB10:%.*]] = sub nsw i32 [[TMP69]], 0
 // CHECK11-NEXT:    [[DIV11:%.*]] = sdiv i32 [[SUB10]], 1
 // CHECK11-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[DIV11]], 1
 // CHECK11-NEXT:    store i32 [[SUB12]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[TMP62:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[ADD13:%.*]] = add nsw i32 [[TMP62]], 1
-// CHECK11-NEXT:    [[TMP63:%.*]] = zext i32 [[ADD13]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP63]])
-// CHECK11-NEXT:    [[TMP64:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50.region_id, i32 4, i8** [[TMP58]], i8** [[TMP59]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP65:%.*]] = icmp ne i32 [[TMP64]], 0
-// CHECK11-NEXT:    br i1 [[TMP65]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK11:       omp_offload.failed14:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50(i32 [[TMP34]], i32* [[TMP35]], i32* [[TMP36]], i32* [[TMP37]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK11:       omp_offload.cont15:
-// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK11-NEXT:    store i32 [[TMP66]], i32* [[CH_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP67:%.*]] = load i32, i32* [[CH_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP68:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP68]], i32* [[N_CASTED16]], align 4
-// CHECK11-NEXT:    [[TMP69:%.*]] = load i32, i32* [[N_CASTED16]], align 4
-// CHECK11-NEXT:    [[TMP70:%.*]] = load i32*, i32** [[A]], align 4
-// CHECK11-NEXT:    [[TMP71:%.*]] = load i32*, i32** [[B]], align 4
-// CHECK11-NEXT:    [[TMP72:%.*]] = load i32*, i32** [[C]], align 4
-// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP67]], i32* [[TMP74]], align 4
-// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP67]], i32* [[TMP76]], align 4
-// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP77]], align 4
-// CHECK11-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP69]], i32* [[TMP79]], align 4
-// CHECK11-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP69]], i32* [[TMP81]], align 4
-// CHECK11-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP82]], align 4
-// CHECK11-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP70]], i32** [[TMP84]], align 4
-// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP70]], i32** [[TMP86]], align 4
-// CHECK11-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP87]], align 4
-// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP71]], i32** [[TMP89]], align 4
-// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP71]], i32** [[TMP91]], align 4
-// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP92]], align 4
-// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP72]], i32** [[TMP94]], align 4
-// CHECK11-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP72]], i32** [[TMP96]], align 4
-// CHECK11-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 4
-// CHECK11-NEXT:    store i8* null, i8** [[TMP97]], align 4
-// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP100]], i32* [[DOTCAPTURE_EXPR_21]], align 4
-// CHECK11-NEXT:    [[TMP101:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_21]], align 4
-// CHECK11-NEXT:    [[SUB23:%.*]] = sub nsw i32 [[TMP101]], 0
-// CHECK11-NEXT:    [[DIV24:%.*]] = sdiv i32 [[SUB23]], 1
-// CHECK11-NEXT:    [[SUB25:%.*]] = sub nsw i32 [[DIV24]], 1
-// CHECK11-NEXT:    store i32 [[SUB25]], i32* [[DOTCAPTURE_EXPR_22]], align 4
-// CHECK11-NEXT:    [[TMP102:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
-// CHECK11-NEXT:    [[ADD26:%.*]] = add nsw i32 [[TMP102]], 1
-// CHECK11-NEXT:    [[TMP103:%.*]] = zext i32 [[ADD26]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP103]])
-// CHECK11-NEXT:    [[TMP104:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l58.region_id, i32 5, i8** [[TMP98]], i8** [[TMP99]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP105:%.*]] = icmp ne i32 [[TMP104]], 0
-// CHECK11-NEXT:    br i1 [[TMP105]], label [[OMP_OFFLOAD_FAILED27:%.*]], label [[OMP_OFFLOAD_CONT28:%.*]]
-// CHECK11:       omp_offload.failed27:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l58(i32 [[TMP67]], i32 [[TMP69]], i32* [[TMP70]], i32* [[TMP71]], i32* [[TMP72]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT28]]
-// CHECK11:       omp_offload.cont28:
-// CHECK11-NEXT:    [[TMP106:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP106]], i32* [[N_CASTED29]], align 4
-// CHECK11-NEXT:    [[TMP107:%.*]] = load i32, i32* [[N_CASTED29]], align 4
-// CHECK11-NEXT:    [[TMP108:%.*]] = load i32*, i32** [[A]], align 4
-// CHECK11-NEXT:    [[TMP109:%.*]] = load i32*, i32** [[B]], align 4
-// CHECK11-NEXT:    [[TMP110:%.*]] = load i32*, i32** [[C]], align 4
-// CHECK11-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP107]], i32* [[TMP112]], align 4
-// CHECK11-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP107]], i32* [[TMP114]], align 4
-// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP115]], align 4
-// CHECK11-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP108]], i32** [[TMP117]], align 4
-// CHECK11-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP119:%.*]] = bitcast i8** [[TMP118]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP108]], i32** [[TMP119]], align 4
-// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP120]], align 4
-// CHECK11-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP109]], i32** [[TMP122]], align 4
-// CHECK11-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP109]], i32** [[TMP124]], align 4
-// CHECK11-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP125]], align 4
-// CHECK11-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP110]], i32** [[TMP127]], align 4
-// CHECK11-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP110]], i32** [[TMP129]], align 4
-// CHECK11-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP130]], align 4
-// CHECK11-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP133:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP133]], i32* [[DOTCAPTURE_EXPR_34]], align 4
-// CHECK11-NEXT:    [[TMP134:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_34]], align 4
-// CHECK11-NEXT:    [[SUB36:%.*]] = sub nsw i32 [[TMP134]], 0
-// CHECK11-NEXT:    [[DIV37:%.*]] = sdiv i32 [[SUB36]], 1
-// CHECK11-NEXT:    [[SUB38:%.*]] = sub nsw i32 [[DIV37]], 1
-// CHECK11-NEXT:    store i32 [[SUB38]], i32* [[DOTCAPTURE_EXPR_35]], align 4
-// CHECK11-NEXT:    [[TMP135:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_35]], align 4
-// CHECK11-NEXT:    [[ADD39:%.*]] = add nsw i32 [[TMP135]], 1
-// CHECK11-NEXT:    [[TMP136:%.*]] = zext i32 [[ADD39]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP136]])
-// CHECK11-NEXT:    [[TMP137:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l66.region_id, i32 4, i8** [[TMP131]], i8** [[TMP132]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.40, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.41, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP138:%.*]] = icmp ne i32 [[TMP137]], 0
-// CHECK11-NEXT:    br i1 [[TMP138]], label [[OMP_OFFLOAD_FAILED40:%.*]], label [[OMP_OFFLOAD_CONT41:%.*]]
-// CHECK11:       omp_offload.failed40:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l66(i32 [[TMP107]], i32* [[TMP108]], i32* [[TMP109]], i32* [[TMP110]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT41]]
-// CHECK11:       omp_offload.cont41:
-// CHECK11-NEXT:    [[TMP139:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK11-NEXT:    store i32 [[TMP139]], i32* [[CH_CASTED42]], align 4
-// CHECK11-NEXT:    [[TMP140:%.*]] = load i32, i32* [[CH_CASTED42]], align 4
-// CHECK11-NEXT:    [[TMP141:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP141]], i32* [[N_CASTED43]], align 4
-// CHECK11-NEXT:    [[TMP142:%.*]] = load i32, i32* [[N_CASTED43]], align 4
-// CHECK11-NEXT:    [[TMP143:%.*]] = load i32*, i32** [[A]], align 4
-// CHECK11-NEXT:    [[TMP144:%.*]] = load i32*, i32** [[B]], align 4
-// CHECK11-NEXT:    [[TMP145:%.*]] = load i32*, i32** [[C]], align 4
-// CHECK11-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP140]], i32* [[TMP147]], align 4
-// CHECK11-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP140]], i32* [[TMP149]], align 4
-// CHECK11-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP150]], align 4
-// CHECK11-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP142]], i32* [[TMP152]], align 4
-// CHECK11-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP142]], i32* [[TMP154]], align 4
-// CHECK11-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP155]], align 4
-// CHECK11-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP143]], i32** [[TMP157]], align 4
-// CHECK11-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP159:%.*]] = bitcast i8** [[TMP158]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP143]], i32** [[TMP159]], align 4
-// CHECK11-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP160]], align 4
-// CHECK11-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP162:%.*]] = bitcast i8** [[TMP161]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP144]], i32** [[TMP162]], align 4
-// CHECK11-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP164:%.*]] = bitcast i8** [[TMP163]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP144]], i32** [[TMP164]], align 4
-// CHECK11-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP165]], align 4
-// CHECK11-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP145]], i32** [[TMP167]], align 4
-// CHECK11-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 4
-// CHECK11-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP145]], i32** [[TMP169]], align 4
-// CHECK11-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i32 0, i32 4
-// CHECK11-NEXT:    store i8* null, i8** [[TMP170]], align 4
-// CHECK11-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP70:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[ADD13:%.*]] = add nsw i32 [[TMP70]], 1
+// CHECK11-NEXT:    [[TMP71:%.*]] = zext i32 [[ADD13]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP71]])
+// CHECK11-NEXT:    [[KERNEL_ARGS14:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP72]], align 4
+// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP73]], align 4
+// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP66]], i8*** [[TMP74]], align 4
+// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP67]], i8*** [[TMP75]], align 4
+// CHECK11-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP76]], align 4
+// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP77]], align 4
+// CHECK11-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP78]], align 4
+// CHECK11-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP79]], align 4
+// CHECK11-NEXT:    [[TMP80:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS14]])
+// CHECK11-NEXT:    [[TMP81:%.*]] = icmp ne i32 [[TMP80]], 0
+// CHECK11-NEXT:    br i1 [[TMP81]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
+// CHECK11:       omp_offload.failed15:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50(i32 [[TMP42]], i32* [[TMP43]], i32* [[TMP44]], i32* [[TMP45]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
+// CHECK11:       omp_offload.cont16:
+// CHECK11-NEXT:    [[TMP82:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK11-NEXT:    store i32 [[TMP82]], i32* [[CH_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP83:%.*]] = load i32, i32* [[CH_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP84:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP84]], i32* [[N_CASTED17]], align 4
+// CHECK11-NEXT:    [[TMP85:%.*]] = load i32, i32* [[N_CASTED17]], align 4
+// CHECK11-NEXT:    [[TMP86:%.*]] = load i32*, i32** [[A]], align 4
+// CHECK11-NEXT:    [[TMP87:%.*]] = load i32*, i32** [[B]], align 4
+// CHECK11-NEXT:    [[TMP88:%.*]] = load i32*, i32** [[C]], align 4
+// CHECK11-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP83]], i32* [[TMP90]], align 4
+// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP83]], i32* [[TMP92]], align 4
+// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP93]], align 4
+// CHECK11-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP85]], i32* [[TMP95]], align 4
+// CHECK11-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP85]], i32* [[TMP97]], align 4
+// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP98]], align 4
+// CHECK11-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP100:%.*]] = bitcast i8** [[TMP99]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP86]], i32** [[TMP100]], align 4
+// CHECK11-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP86]], i32** [[TMP102]], align 4
+// CHECK11-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP103]], align 4
+// CHECK11-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP87]], i32** [[TMP105]], align 4
+// CHECK11-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP87]], i32** [[TMP107]], align 4
+// CHECK11-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP108]], align 4
+// CHECK11-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP88]], i32** [[TMP110]], align 4
+// CHECK11-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP88]], i32** [[TMP112]], align 4
+// CHECK11-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 4
+// CHECK11-NEXT:    store i8* null, i8** [[TMP113]], align 4
+// CHECK11-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP116:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP116]], i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK11-NEXT:    [[TMP117:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK11-NEXT:    [[SUB24:%.*]] = sub nsw i32 [[TMP117]], 0
+// CHECK11-NEXT:    [[DIV25:%.*]] = sdiv i32 [[SUB24]], 1
+// CHECK11-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[DIV25]], 1
+// CHECK11-NEXT:    store i32 [[SUB26]], i32* [[DOTCAPTURE_EXPR_23]], align 4
+// CHECK11-NEXT:    [[TMP118:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_23]], align 4
+// CHECK11-NEXT:    [[ADD27:%.*]] = add nsw i32 [[TMP118]], 1
+// CHECK11-NEXT:    [[TMP119:%.*]] = zext i32 [[ADD27]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP119]])
+// CHECK11-NEXT:    [[KERNEL_ARGS28:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP120]], align 4
+// CHECK11-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP121]], align 4
+// CHECK11-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP114]], i8*** [[TMP122]], align 4
+// CHECK11-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP115]], i8*** [[TMP123]], align 4
+// CHECK11-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP124]], align 4
+// CHECK11-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP125]], align 4
+// CHECK11-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP126]], align 4
+// CHECK11-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP127]], align 4
+// CHECK11-NEXT:    [[TMP128:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l58.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]])
+// CHECK11-NEXT:    [[TMP129:%.*]] = icmp ne i32 [[TMP128]], 0
+// CHECK11-NEXT:    br i1 [[TMP129]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
+// CHECK11:       omp_offload.failed29:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l58(i32 [[TMP83]], i32 [[TMP85]], i32* [[TMP86]], i32* [[TMP87]], i32* [[TMP88]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
+// CHECK11:       omp_offload.cont30:
+// CHECK11-NEXT:    [[TMP130:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP130]], i32* [[N_CASTED31]], align 4
+// CHECK11-NEXT:    [[TMP131:%.*]] = load i32, i32* [[N_CASTED31]], align 4
+// CHECK11-NEXT:    [[TMP132:%.*]] = load i32*, i32** [[A]], align 4
+// CHECK11-NEXT:    [[TMP133:%.*]] = load i32*, i32** [[B]], align 4
+// CHECK11-NEXT:    [[TMP134:%.*]] = load i32*, i32** [[C]], align 4
+// CHECK11-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP131]], i32* [[TMP136]], align 4
+// CHECK11-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP131]], i32* [[TMP138]], align 4
+// CHECK11-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP139]], align 4
+// CHECK11-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP132]], i32** [[TMP141]], align 4
+// CHECK11-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP132]], i32** [[TMP143]], align 4
+// CHECK11-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP144]], align 4
+// CHECK11-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP133]], i32** [[TMP146]], align 4
+// CHECK11-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP133]], i32** [[TMP148]], align 4
+// CHECK11-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP149]], align 4
+// CHECK11-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP134]], i32** [[TMP151]], align 4
+// CHECK11-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP134]], i32** [[TMP153]], align 4
+// CHECK11-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP154]], align 4
+// CHECK11-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP157:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP157]], i32* [[DOTCAPTURE_EXPR_36]], align 4
+// CHECK11-NEXT:    [[TMP158:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_36]], align 4
+// CHECK11-NEXT:    [[SUB38:%.*]] = sub nsw i32 [[TMP158]], 0
+// CHECK11-NEXT:    [[DIV39:%.*]] = sdiv i32 [[SUB38]], 1
+// CHECK11-NEXT:    [[SUB40:%.*]] = sub nsw i32 [[DIV39]], 1
+// CHECK11-NEXT:    store i32 [[SUB40]], i32* [[DOTCAPTURE_EXPR_37]], align 4
+// CHECK11-NEXT:    [[TMP159:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_37]], align 4
+// CHECK11-NEXT:    [[ADD41:%.*]] = add nsw i32 [[TMP159]], 1
+// CHECK11-NEXT:    [[TMP160:%.*]] = zext i32 [[ADD41]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP160]])
+// CHECK11-NEXT:    [[KERNEL_ARGS42:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP161]], align 4
+// CHECK11-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP162]], align 4
+// CHECK11-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP155]], i8*** [[TMP163]], align 4
+// CHECK11-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP156]], i8*** [[TMP164]], align 4
+// CHECK11-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.40, i32 0, i32 0), i64** [[TMP165]], align 4
+// CHECK11-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.41, i32 0, i32 0), i64** [[TMP166]], align 4
+// CHECK11-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP167]], align 4
+// CHECK11-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP168]], align 4
+// CHECK11-NEXT:    [[TMP169:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l66.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS42]])
+// CHECK11-NEXT:    [[TMP170:%.*]] = icmp ne i32 [[TMP169]], 0
+// CHECK11-NEXT:    br i1 [[TMP170]], label [[OMP_OFFLOAD_FAILED43:%.*]], label [[OMP_OFFLOAD_CONT44:%.*]]
+// CHECK11:       omp_offload.failed43:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l66(i32 [[TMP131]], i32* [[TMP132]], i32* [[TMP133]], i32* [[TMP134]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT44]]
+// CHECK11:       omp_offload.cont44:
+// CHECK11-NEXT:    [[TMP171:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK11-NEXT:    store i32 [[TMP171]], i32* [[CH_CASTED45]], align 4
+// CHECK11-NEXT:    [[TMP172:%.*]] = load i32, i32* [[CH_CASTED45]], align 4
 // CHECK11-NEXT:    [[TMP173:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP173]], i32* [[DOTCAPTURE_EXPR_48]], align 4
-// CHECK11-NEXT:    [[TMP174:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_48]], align 4
-// CHECK11-NEXT:    [[SUB50:%.*]] = sub nsw i32 [[TMP174]], 0
-// CHECK11-NEXT:    [[DIV51:%.*]] = sdiv i32 [[SUB50]], 1
-// CHECK11-NEXT:    [[SUB52:%.*]] = sub nsw i32 [[DIV51]], 1
-// CHECK11-NEXT:    store i32 [[SUB52]], i32* [[DOTCAPTURE_EXPR_49]], align 4
-// CHECK11-NEXT:    [[TMP175:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_49]], align 4
-// CHECK11-NEXT:    [[ADD53:%.*]] = add nsw i32 [[TMP175]], 1
-// CHECK11-NEXT:    [[TMP176:%.*]] = zext i32 [[ADD53]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP176]])
-// CHECK11-NEXT:    [[TMP177:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l74.region_id, i32 5, i8** [[TMP171]], i8** [[TMP172]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.44, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.45, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP178:%.*]] = icmp ne i32 [[TMP177]], 0
-// CHECK11-NEXT:    br i1 [[TMP178]], label [[OMP_OFFLOAD_FAILED54:%.*]], label [[OMP_OFFLOAD_CONT55:%.*]]
-// CHECK11:       omp_offload.failed54:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l74(i32 [[TMP140]], i32 [[TMP142]], i32* [[TMP143]], i32* [[TMP144]], i32* [[TMP145]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT55]]
-// CHECK11:       omp_offload.cont55:
-// CHECK11-NEXT:    [[TMP179:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP179]], i32* [[N_CASTED56]], align 4
-// CHECK11-NEXT:    [[TMP180:%.*]] = load i32, i32* [[N_CASTED56]], align 4
-// CHECK11-NEXT:    [[TMP181:%.*]] = load i32*, i32** [[A]], align 4
-// CHECK11-NEXT:    [[TMP182:%.*]] = load i32*, i32** [[B]], align 4
-// CHECK11-NEXT:    [[TMP183:%.*]] = load i32*, i32** [[C]], align 4
-// CHECK11-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP185:%.*]] = bitcast i8** [[TMP184]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP180]], i32* [[TMP185]], align 4
-// CHECK11-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP187:%.*]] = bitcast i8** [[TMP186]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP180]], i32* [[TMP187]], align 4
-// CHECK11-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP188]], align 4
-// CHECK11-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP190:%.*]] = bitcast i8** [[TMP189]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP181]], i32** [[TMP190]], align 4
-// CHECK11-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP192:%.*]] = bitcast i8** [[TMP191]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP181]], i32** [[TMP192]], align 4
-// CHECK11-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP193]], align 4
-// CHECK11-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP195:%.*]] = bitcast i8** [[TMP194]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP182]], i32** [[TMP195]], align 4
-// CHECK11-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP197:%.*]] = bitcast i8** [[TMP196]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP182]], i32** [[TMP197]], align 4
-// CHECK11-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP198]], align 4
-// CHECK11-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP200:%.*]] = bitcast i8** [[TMP199]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP183]], i32** [[TMP200]], align 4
-// CHECK11-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP202:%.*]] = bitcast i8** [[TMP201]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP183]], i32** [[TMP202]], align 4
-// CHECK11-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS59]], i32 0, i32 3
-// CHECK11-NEXT:    store i8* null, i8** [[TMP203]], align 4
-// CHECK11-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS57]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS58]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP206:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP206]], i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK11-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK11-NEXT:    [[SUB63:%.*]] = sub nsw i32 [[TMP207]], 0
-// CHECK11-NEXT:    [[DIV64:%.*]] = sdiv i32 [[SUB63]], 1
-// CHECK11-NEXT:    [[SUB65:%.*]] = sub nsw i32 [[DIV64]], 1
-// CHECK11-NEXT:    store i32 [[SUB65]], i32* [[DOTCAPTURE_EXPR_62]], align 4
-// CHECK11-NEXT:    [[TMP208:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_62]], align 4
-// CHECK11-NEXT:    [[ADD66:%.*]] = add nsw i32 [[TMP208]], 1
-// CHECK11-NEXT:    [[TMP209:%.*]] = zext i32 [[ADD66]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP209]])
-// CHECK11-NEXT:    [[TMP210:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l82.region_id, i32 4, i8** [[TMP204]], i8** [[TMP205]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.48, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.49, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP211:%.*]] = icmp ne i32 [[TMP210]], 0
-// CHECK11-NEXT:    br i1 [[TMP211]], label [[OMP_OFFLOAD_FAILED67:%.*]], label [[OMP_OFFLOAD_CONT68:%.*]]
-// CHECK11:       omp_offload.failed67:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l82(i32 [[TMP180]], i32* [[TMP181]], i32* [[TMP182]], i32* [[TMP183]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT68]]
-// CHECK11:       omp_offload.cont68:
-// CHECK11-NEXT:    [[TMP212:%.*]] = load i32, i32* [[CH]], align 4
-// CHECK11-NEXT:    store i32 [[TMP212]], i32* [[CH_CASTED69]], align 4
-// CHECK11-NEXT:    [[TMP213:%.*]] = load i32, i32* [[CH_CASTED69]], align 4
-// CHECK11-NEXT:    [[TMP214:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP214]], i32* [[N_CASTED70]], align 4
-// CHECK11-NEXT:    [[TMP215:%.*]] = load i32, i32* [[N_CASTED70]], align 4
-// CHECK11-NEXT:    [[TMP216:%.*]] = load i32*, i32** [[A]], align 4
-// CHECK11-NEXT:    [[TMP217:%.*]] = load i32*, i32** [[B]], align 4
-// CHECK11-NEXT:    [[TMP218:%.*]] = load i32*, i32** [[C]], align 4
-// CHECK11-NEXT:    [[TMP219:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP220:%.*]] = bitcast i8** [[TMP219]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP213]], i32* [[TMP220]], align 4
-// CHECK11-NEXT:    [[TMP221:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP222:%.*]] = bitcast i8** [[TMP221]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP213]], i32* [[TMP222]], align 4
-// CHECK11-NEXT:    [[TMP223:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP223]], align 4
-// CHECK11-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 [[TMP173]], i32* [[N_CASTED46]], align 4
+// CHECK11-NEXT:    [[TMP174:%.*]] = load i32, i32* [[N_CASTED46]], align 4
+// CHECK11-NEXT:    [[TMP175:%.*]] = load i32*, i32** [[A]], align 4
+// CHECK11-NEXT:    [[TMP176:%.*]] = load i32*, i32** [[B]], align 4
+// CHECK11-NEXT:    [[TMP177:%.*]] = load i32*, i32** [[C]], align 4
+// CHECK11-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP172]], i32* [[TMP179]], align 4
+// CHECK11-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP172]], i32* [[TMP181]], align 4
+// CHECK11-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP182]], align 4
+// CHECK11-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP184:%.*]] = bitcast i8** [[TMP183]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP174]], i32* [[TMP184]], align 4
+// CHECK11-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP186:%.*]] = bitcast i8** [[TMP185]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP174]], i32* [[TMP186]], align 4
+// CHECK11-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP187]], align 4
+// CHECK11-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP189:%.*]] = bitcast i8** [[TMP188]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP175]], i32** [[TMP189]], align 4
+// CHECK11-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP191:%.*]] = bitcast i8** [[TMP190]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP175]], i32** [[TMP191]], align 4
+// CHECK11-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP192]], align 4
+// CHECK11-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP194:%.*]] = bitcast i8** [[TMP193]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP176]], i32** [[TMP194]], align 4
+// CHECK11-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP196:%.*]] = bitcast i8** [[TMP195]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP176]], i32** [[TMP196]], align 4
+// CHECK11-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP197]], align 4
+// CHECK11-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP199:%.*]] = bitcast i8** [[TMP198]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP177]], i32** [[TMP199]], align 4
+// CHECK11-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP201:%.*]] = bitcast i8** [[TMP200]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP177]], i32** [[TMP201]], align 4
+// CHECK11-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 4
+// CHECK11-NEXT:    store i8* null, i8** [[TMP202]], align 4
+// CHECK11-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP205:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP205]], i32* [[DOTCAPTURE_EXPR_51]], align 4
+// CHECK11-NEXT:    [[TMP206:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_51]], align 4
+// CHECK11-NEXT:    [[SUB53:%.*]] = sub nsw i32 [[TMP206]], 0
+// CHECK11-NEXT:    [[DIV54:%.*]] = sdiv i32 [[SUB53]], 1
+// CHECK11-NEXT:    [[SUB55:%.*]] = sub nsw i32 [[DIV54]], 1
+// CHECK11-NEXT:    store i32 [[SUB55]], i32* [[DOTCAPTURE_EXPR_52]], align 4
+// CHECK11-NEXT:    [[TMP207:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_52]], align 4
+// CHECK11-NEXT:    [[ADD56:%.*]] = add nsw i32 [[TMP207]], 1
+// CHECK11-NEXT:    [[TMP208:%.*]] = zext i32 [[ADD56]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP208]])
+// CHECK11-NEXT:    [[KERNEL_ARGS57:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP209:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP209]], align 4
+// CHECK11-NEXT:    [[TMP210:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP210]], align 4
+// CHECK11-NEXT:    [[TMP211:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP203]], i8*** [[TMP211]], align 4
+// CHECK11-NEXT:    [[TMP212:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP204]], i8*** [[TMP212]], align 4
+// CHECK11-NEXT:    [[TMP213:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.44, i32 0, i32 0), i64** [[TMP213]], align 4
+// CHECK11-NEXT:    [[TMP214:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.45, i32 0, i32 0), i64** [[TMP214]], align 4
+// CHECK11-NEXT:    [[TMP215:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP215]], align 4
+// CHECK11-NEXT:    [[TMP216:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP216]], align 4
+// CHECK11-NEXT:    [[TMP217:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l74.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS57]])
+// CHECK11-NEXT:    [[TMP218:%.*]] = icmp ne i32 [[TMP217]], 0
+// CHECK11-NEXT:    br i1 [[TMP218]], label [[OMP_OFFLOAD_FAILED58:%.*]], label [[OMP_OFFLOAD_CONT59:%.*]]
+// CHECK11:       omp_offload.failed58:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l74(i32 [[TMP172]], i32 [[TMP174]], i32* [[TMP175]], i32* [[TMP176]], i32* [[TMP177]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT59]]
+// CHECK11:       omp_offload.cont59:
+// CHECK11-NEXT:    [[TMP219:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP219]], i32* [[N_CASTED60]], align 4
+// CHECK11-NEXT:    [[TMP220:%.*]] = load i32, i32* [[N_CASTED60]], align 4
+// CHECK11-NEXT:    [[TMP221:%.*]] = load i32*, i32** [[A]], align 4
+// CHECK11-NEXT:    [[TMP222:%.*]] = load i32*, i32** [[B]], align 4
+// CHECK11-NEXT:    [[TMP223:%.*]] = load i32*, i32** [[C]], align 4
+// CHECK11-NEXT:    [[TMP224:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP225:%.*]] = bitcast i8** [[TMP224]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP215]], i32* [[TMP225]], align 4
-// CHECK11-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 [[TMP220]], i32* [[TMP225]], align 4
+// CHECK11-NEXT:    [[TMP226:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP227:%.*]] = bitcast i8** [[TMP226]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP215]], i32* [[TMP227]], align 4
-// CHECK11-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 [[TMP220]], i32* [[TMP227]], align 4
+// CHECK11-NEXT:    [[TMP228:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 0
 // CHECK11-NEXT:    store i8* null, i8** [[TMP228]], align 4
-// CHECK11-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP229:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 1
 // CHECK11-NEXT:    [[TMP230:%.*]] = bitcast i8** [[TMP229]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP216]], i32** [[TMP230]], align 4
-// CHECK11-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 2
+// CHECK11-NEXT:    store i32* [[TMP221]], i32** [[TMP230]], align 4
+// CHECK11-NEXT:    [[TMP231:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 1
 // CHECK11-NEXT:    [[TMP232:%.*]] = bitcast i8** [[TMP231]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP216]], i32** [[TMP232]], align 4
-// CHECK11-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 2
+// CHECK11-NEXT:    store i32* [[TMP221]], i32** [[TMP232]], align 4
+// CHECK11-NEXT:    [[TMP233:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 1
 // CHECK11-NEXT:    store i8* null, i8** [[TMP233]], align 4
-// CHECK11-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP234:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 2
 // CHECK11-NEXT:    [[TMP235:%.*]] = bitcast i8** [[TMP234]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP217]], i32** [[TMP235]], align 4
-// CHECK11-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 3
+// CHECK11-NEXT:    store i32* [[TMP222]], i32** [[TMP235]], align 4
+// CHECK11-NEXT:    [[TMP236:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 2
 // CHECK11-NEXT:    [[TMP237:%.*]] = bitcast i8** [[TMP236]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP217]], i32** [[TMP237]], align 4
-// CHECK11-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 3
+// CHECK11-NEXT:    store i32* [[TMP222]], i32** [[TMP237]], align 4
+// CHECK11-NEXT:    [[TMP238:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 2
 // CHECK11-NEXT:    store i8* null, i8** [[TMP238]], align 4
-// CHECK11-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP239:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 3
 // CHECK11-NEXT:    [[TMP240:%.*]] = bitcast i8** [[TMP239]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP218]], i32** [[TMP240]], align 4
-// CHECK11-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 4
+// CHECK11-NEXT:    store i32* [[TMP223]], i32** [[TMP240]], align 4
+// CHECK11-NEXT:    [[TMP241:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 3
 // CHECK11-NEXT:    [[TMP242:%.*]] = bitcast i8** [[TMP241]] to i32**
-// CHECK11-NEXT:    store i32* [[TMP218]], i32** [[TMP242]], align 4
-// CHECK11-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS73]], i32 0, i32 4
+// CHECK11-NEXT:    store i32* [[TMP223]], i32** [[TMP242]], align 4
+// CHECK11-NEXT:    [[TMP243:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS63]], i32 0, i32 3
 // CHECK11-NEXT:    store i8* null, i8** [[TMP243]], align 4
-// CHECK11-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS71]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS72]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP244:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS61]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP245:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS62]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP246:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_75]], align 4
-// CHECK11-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_75]], align 4
-// CHECK11-NEXT:    [[SUB77:%.*]] = sub nsw i32 [[TMP247]], 0
-// CHECK11-NEXT:    [[DIV78:%.*]] = sdiv i32 [[SUB77]], 1
-// CHECK11-NEXT:    [[SUB79:%.*]] = sub nsw i32 [[DIV78]], 1
-// CHECK11-NEXT:    store i32 [[SUB79]], i32* [[DOTCAPTURE_EXPR_76]], align 4
-// CHECK11-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_76]], align 4
-// CHECK11-NEXT:    [[ADD80:%.*]] = add nsw i32 [[TMP248]], 1
-// CHECK11-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD80]] to i64
+// CHECK11-NEXT:    store i32 [[TMP246]], i32* [[DOTCAPTURE_EXPR_65]], align 4
+// CHECK11-NEXT:    [[TMP247:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_65]], align 4
+// CHECK11-NEXT:    [[SUB67:%.*]] = sub nsw i32 [[TMP247]], 0
+// CHECK11-NEXT:    [[DIV68:%.*]] = sdiv i32 [[SUB67]], 1
+// CHECK11-NEXT:    [[SUB69:%.*]] = sub nsw i32 [[DIV68]], 1
+// CHECK11-NEXT:    store i32 [[SUB69]], i32* [[DOTCAPTURE_EXPR_66]], align 4
+// CHECK11-NEXT:    [[TMP248:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_66]], align 4
+// CHECK11-NEXT:    [[ADD70:%.*]] = add nsw i32 [[TMP248]], 1
+// CHECK11-NEXT:    [[TMP249:%.*]] = zext i32 [[ADD70]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP249]])
-// CHECK11-NEXT:    [[TMP250:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l90.region_id, i32 5, i8** [[TMP244]], i8** [[TMP245]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.52, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.53, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP251:%.*]] = icmp ne i32 [[TMP250]], 0
-// CHECK11-NEXT:    br i1 [[TMP251]], label [[OMP_OFFLOAD_FAILED81:%.*]], label [[OMP_OFFLOAD_CONT82:%.*]]
-// CHECK11:       omp_offload.failed81:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l90(i32 [[TMP213]], i32 [[TMP215]], i32* [[TMP216]], i32* [[TMP217]], i32* [[TMP218]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT82]]
-// CHECK11:       omp_offload.cont82:
+// CHECK11-NEXT:    [[KERNEL_ARGS71:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP250:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP250]], align 4
+// CHECK11-NEXT:    [[TMP251:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP251]], align 4
+// CHECK11-NEXT:    [[TMP252:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP244]], i8*** [[TMP252]], align 4
+// CHECK11-NEXT:    [[TMP253:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP245]], i8*** [[TMP253]], align 4
+// CHECK11-NEXT:    [[TMP254:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.48, i32 0, i32 0), i64** [[TMP254]], align 4
+// CHECK11-NEXT:    [[TMP255:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.49, i32 0, i32 0), i64** [[TMP255]], align 4
+// CHECK11-NEXT:    [[TMP256:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP256]], align 4
+// CHECK11-NEXT:    [[TMP257:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP257]], align 4
+// CHECK11-NEXT:    [[TMP258:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l82.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS71]])
+// CHECK11-NEXT:    [[TMP259:%.*]] = icmp ne i32 [[TMP258]], 0
+// CHECK11-NEXT:    br i1 [[TMP259]], label [[OMP_OFFLOAD_FAILED72:%.*]], label [[OMP_OFFLOAD_CONT73:%.*]]
+// CHECK11:       omp_offload.failed72:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l82(i32 [[TMP220]], i32* [[TMP221]], i32* [[TMP222]], i32* [[TMP223]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT73]]
+// CHECK11:       omp_offload.cont73:
+// CHECK11-NEXT:    [[TMP260:%.*]] = load i32, i32* [[CH]], align 4
+// CHECK11-NEXT:    store i32 [[TMP260]], i32* [[CH_CASTED74]], align 4
+// CHECK11-NEXT:    [[TMP261:%.*]] = load i32, i32* [[CH_CASTED74]], align 4
+// CHECK11-NEXT:    [[TMP262:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP262]], i32* [[N_CASTED75]], align 4
+// CHECK11-NEXT:    [[TMP263:%.*]] = load i32, i32* [[N_CASTED75]], align 4
+// CHECK11-NEXT:    [[TMP264:%.*]] = load i32*, i32** [[A]], align 4
+// CHECK11-NEXT:    [[TMP265:%.*]] = load i32*, i32** [[B]], align 4
+// CHECK11-NEXT:    [[TMP266:%.*]] = load i32*, i32** [[C]], align 4
+// CHECK11-NEXT:    [[TMP267:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP268:%.*]] = bitcast i8** [[TMP267]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP261]], i32* [[TMP268]], align 4
+// CHECK11-NEXT:    [[TMP269:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP270:%.*]] = bitcast i8** [[TMP269]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP261]], i32* [[TMP270]], align 4
+// CHECK11-NEXT:    [[TMP271:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP271]], align 4
+// CHECK11-NEXT:    [[TMP272:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP273:%.*]] = bitcast i8** [[TMP272]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP263]], i32* [[TMP273]], align 4
+// CHECK11-NEXT:    [[TMP274:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP275:%.*]] = bitcast i8** [[TMP274]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP263]], i32* [[TMP275]], align 4
+// CHECK11-NEXT:    [[TMP276:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP276]], align 4
+// CHECK11-NEXT:    [[TMP277:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP278:%.*]] = bitcast i8** [[TMP277]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP264]], i32** [[TMP278]], align 4
+// CHECK11-NEXT:    [[TMP279:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP280:%.*]] = bitcast i8** [[TMP279]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP264]], i32** [[TMP280]], align 4
+// CHECK11-NEXT:    [[TMP281:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP281]], align 4
+// CHECK11-NEXT:    [[TMP282:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP283:%.*]] = bitcast i8** [[TMP282]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP265]], i32** [[TMP283]], align 4
+// CHECK11-NEXT:    [[TMP284:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP285:%.*]] = bitcast i8** [[TMP284]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP265]], i32** [[TMP285]], align 4
+// CHECK11-NEXT:    [[TMP286:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP286]], align 4
+// CHECK11-NEXT:    [[TMP287:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP288:%.*]] = bitcast i8** [[TMP287]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP266]], i32** [[TMP288]], align 4
+// CHECK11-NEXT:    [[TMP289:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 4
+// CHECK11-NEXT:    [[TMP290:%.*]] = bitcast i8** [[TMP289]] to i32**
+// CHECK11-NEXT:    store i32* [[TMP266]], i32** [[TMP290]], align 4
+// CHECK11-NEXT:    [[TMP291:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS78]], i32 0, i32 4
+// CHECK11-NEXT:    store i8* null, i8** [[TMP291]], align 4
+// CHECK11-NEXT:    [[TMP292:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS76]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP293:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS77]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP294:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP294]], i32* [[DOTCAPTURE_EXPR_80]], align 4
+// CHECK11-NEXT:    [[TMP295:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_80]], align 4
+// CHECK11-NEXT:    [[SUB82:%.*]] = sub nsw i32 [[TMP295]], 0
+// CHECK11-NEXT:    [[DIV83:%.*]] = sdiv i32 [[SUB82]], 1
+// CHECK11-NEXT:    [[SUB84:%.*]] = sub nsw i32 [[DIV83]], 1
+// CHECK11-NEXT:    store i32 [[SUB84]], i32* [[DOTCAPTURE_EXPR_81]], align 4
+// CHECK11-NEXT:    [[TMP296:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_81]], align 4
+// CHECK11-NEXT:    [[ADD85:%.*]] = add nsw i32 [[TMP296]], 1
+// CHECK11-NEXT:    [[TMP297:%.*]] = zext i32 [[ADD85]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP297]])
+// CHECK11-NEXT:    [[KERNEL_ARGS86:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP298:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP298]], align 4
+// CHECK11-NEXT:    [[TMP299:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP299]], align 4
+// CHECK11-NEXT:    [[TMP300:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP292]], i8*** [[TMP300]], align 4
+// CHECK11-NEXT:    [[TMP301:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP293]], i8*** [[TMP301]], align 4
+// CHECK11-NEXT:    [[TMP302:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.52, i32 0, i32 0), i64** [[TMP302]], align 4
+// CHECK11-NEXT:    [[TMP303:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.53, i32 0, i32 0), i64** [[TMP303]], align 4
+// CHECK11-NEXT:    [[TMP304:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP304]], align 4
+// CHECK11-NEXT:    [[TMP305:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP305]], align 4
+// CHECK11-NEXT:    [[TMP306:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l90.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS86]])
+// CHECK11-NEXT:    [[TMP307:%.*]] = icmp ne i32 [[TMP306]], 0
+// CHECK11-NEXT:    br i1 [[TMP307]], label [[OMP_OFFLOAD_FAILED87:%.*]], label [[OMP_OFFLOAD_CONT88:%.*]]
+// CHECK11:       omp_offload.failed87:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l90(i32 [[TMP261]], i32 [[TMP263]], i32* [[TMP264]], i32* [[TMP265]], i32* [[TMP266]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT88]]
+// CHECK11:       omp_offload.cont88:
 // CHECK11-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/distribute_parallel_for_simd_firstprivate_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_simd_firstprivate_codegen.cpp
index f58863b..e211678 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_simd_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_simd_firstprivate_codegen.cpp
@@ -931,9 +931,26 @@ int main() {
 // CHECK8-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK8-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK8-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK8-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l138.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK8-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK8-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK8-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK8-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK8-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK8-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK8-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK8-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK8-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 8
+// CHECK8-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK8-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 8
+// CHECK8-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK8-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK8-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK8-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK8-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK8-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK8-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK8-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK8-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l138.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK8-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK8-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK8:       omp_offload.failed:
 // CHECK8-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l138(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK8-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -941,18 +958,18 @@ int main() {
 // CHECK8-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK8-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK8-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK8-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK8-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK8-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK8:       arraydestroy.body:
-// CHECK8-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK8-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK8-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK8-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK8-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK8-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK8:       arraydestroy.done3:
 // CHECK8-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK8-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK8-NEXT:    ret i32 [[TMP39]]
+// CHECK8-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK8-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK8-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1377,27 +1394,44 @@ int main() {
 // CHECK8-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK8-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK8-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK8-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l48.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK8-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK8-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK8-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK8-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK8-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK8-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK8-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK8-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK8-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK8-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK8-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK8-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK8-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK8-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK8-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK8-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK8-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK8-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK8-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK8-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK8-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK8-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK8:       omp_offload.failed:
 // CHECK8-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l48(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK8-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK8:       omp_offload.cont:
 // CHECK8-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK8-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK8-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK8-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK8-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK8:       arraydestroy.body:
-// CHECK8-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK8-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK8-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK8-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK8-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK8-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK8:       arraydestroy.done2:
 // CHECK8-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK8-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK8-NEXT:    ret i32 [[TMP32]]
+// CHECK8-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK8-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK8-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1895,9 +1929,26 @@ int main() {
 // CHECK10-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK10-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK10-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK10-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l138.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK10-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK10-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK10-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK10-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK10-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK10-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK10-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK10-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK10-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 4
+// CHECK10-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK10-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 4
+// CHECK10-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK10-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK10-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK10-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK10-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK10-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK10-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK10-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK10-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l138.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK10-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK10-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK10:       omp_offload.failed:
 // CHECK10-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l138(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK10-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1905,18 +1956,18 @@ int main() {
 // CHECK10-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK10-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK10-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK10-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK10-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK10-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK10:       arraydestroy.body:
-// CHECK10-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK10-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK10-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK10-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK10-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK10-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK10:       arraydestroy.done2:
 // CHECK10-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK10-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK10-NEXT:    ret i32 [[TMP39]]
+// CHECK10-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK10-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK10-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -2328,27 +2379,44 @@ int main() {
 // CHECK10-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK10-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK10-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK10-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l48.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK10-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK10-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK10-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK10-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK10-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK10-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK10-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK10-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK10-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK10-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK10-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK10-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK10-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK10-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK10-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK10-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK10-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK10-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK10-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK10-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK10-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK10-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK10:       omp_offload.failed:
 // CHECK10-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l48(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK10-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK10:       omp_offload.cont:
 // CHECK10-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK10-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK10-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK10-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK10-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK10:       arraydestroy.body:
-// CHECK10-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK10-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK10-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK10-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK10-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK10-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK10:       arraydestroy.done2:
 // CHECK10-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK10-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK10-NEXT:    ret i32 [[TMP32]]
+// CHECK10-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK10-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK10-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/distribute_parallel_for_simd_if_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_simd_if_codegen.cpp
index cbe3ae5..947da63 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_simd_if_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_simd_if_codegen.cpp
@@ -123,21 +123,55 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43() #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret void
 //
 //
@@ -459,48 +493,99 @@ int main() {
 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
+// CHECK1-NEXT:    [[TMP20:%.*]] = load i32, i32* @Arg, align 4
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK1-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK1:       omp_offload.failed5:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK1:       omp_offload.cont6:
-// CHECK1-NEXT:    [[TMP15:%.*]] = load i32, i32* @Arg, align 4
-// CHECK1-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP15]])
+// CHECK1-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK1-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK1-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK1:       omp_offload.cont8:
+// CHECK1-NEXT:    [[TMP39:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP39]])
 // CHECK1-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -995,46 +1080,97 @@ int main() {
 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l59.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l59.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l59() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l65.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l65.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l65() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
+// CHECK1-NEXT:    [[TMP20:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK1-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK1:       omp_offload.failed5:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK1:       omp_offload.cont6:
+// CHECK1-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK1-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK1-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK1:       omp_offload.cont8:
 // CHECK1-NEXT:    ret i32 0
 //
 //
@@ -1532,21 +1668,55 @@ int main() {
 // CHECK3-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43() #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK3-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK3:       omp_offload.failed2:
+// CHECK3-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK3-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK3-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK3-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK3:       omp_offload.failed3:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48() #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK3:       omp_offload.cont3:
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK3:       omp_offload.cont4:
 // CHECK3-NEXT:    ret void
 //
 //
@@ -1868,48 +2038,99 @@ int main() {
 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
-// CHECK3-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK3-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK3:       omp_offload.failed2:
+// CHECK3-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK3-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK3-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK3-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK3:       omp_offload.failed3:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90() #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK3:       omp_offload.cont3:
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK3:       omp_offload.cont4:
+// CHECK3-NEXT:    [[TMP20:%.*]] = load i32, i32* @Arg, align 4
 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK3-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK3-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK3-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK3-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK3-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK3-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK3:       omp_offload.failed5:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK3:       omp_offload.cont6:
-// CHECK3-NEXT:    [[TMP15:%.*]] = load i32, i32* @Arg, align 4
-// CHECK3-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP15]])
+// CHECK3-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK3-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK3-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK3-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK3:       omp_offload.cont8:
+// CHECK3-NEXT:    [[TMP39:%.*]] = load i32, i32* @Arg, align 4
+// CHECK3-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP39]])
 // CHECK3-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -2643,46 +2864,97 @@ int main() {
 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
-// CHECK3-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l59.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l59.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l59() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l65.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK3-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK3:       omp_offload.failed2:
+// CHECK3-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK3-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l65.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK3-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK3-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK3:       omp_offload.failed3:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l65() #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK3:       omp_offload.cont3:
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK3:       omp_offload.cont4:
+// CHECK3-NEXT:    [[TMP20:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK3-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK3-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK3-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK3-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK3-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK3-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK3:       omp_offload.failed5:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK3:       omp_offload.cont6:
+// CHECK3-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK3-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK3-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK3-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK3:       omp_offload.cont8:
 // CHECK3-NEXT:    ret i32 0
 //
 //
@@ -3756,21 +4028,55 @@ int main() {
 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43() #[[ATTR2:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK9-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK9:       omp_offload.failed2:
+// CHECK9-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK9-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK9-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK9:       omp_offload.failed3:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48() #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK9:       omp_offload.cont3:
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK9:       omp_offload.cont4:
 // CHECK9-NEXT:    ret void
 //
 //
@@ -4092,48 +4398,99 @@ int main() {
 // CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81() #[[ATTR2]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK9-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK9:       omp_offload.failed2:
+// CHECK9-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK9-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK9-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK9:       omp_offload.failed3:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90() #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK9:       omp_offload.cont3:
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK9:       omp_offload.cont4:
+// CHECK9-NEXT:    [[TMP20:%.*]] = load i32, i32* @Arg, align 4
 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK9-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK9-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK9-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK9:       omp_offload.failed5:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK9:       omp_offload.cont6:
-// CHECK9-NEXT:    [[TMP15:%.*]] = load i32, i32* @Arg, align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP15]])
+// CHECK9-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK9-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK9-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK9:       omp_offload.failed7:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK9:       omp_offload.cont8:
+// CHECK9-NEXT:    [[TMP39:%.*]] = load i32, i32* @Arg, align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP39]])
 // CHECK9-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -4628,46 +4985,97 @@ int main() {
 // CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l59.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l59.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l59() #[[ATTR2]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l65.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK9-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK9:       omp_offload.failed2:
+// CHECK9-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l65.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK9-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK9-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK9:       omp_offload.failed3:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l65() #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK9:       omp_offload.cont3:
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK9:       omp_offload.cont4:
+// CHECK9-NEXT:    [[TMP20:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK9-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK9-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK9-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK9:       omp_offload.failed5:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK9:       omp_offload.cont6:
+// CHECK9-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK9-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK9-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK9:       omp_offload.failed7:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK9:       omp_offload.cont8:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -5165,21 +5573,55 @@ int main() {
 // CHECK11-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK11-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK11-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK11-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43() #[[ATTR2:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK11-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK11:       omp_offload.failed2:
+// CHECK11-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK11-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK11-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK11-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK11-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK11:       omp_offload.failed3:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48() #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK11:       omp_offload.cont3:
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK11:       omp_offload.cont4:
 // CHECK11-NEXT:    ret void
 //
 //
@@ -5501,48 +5943,99 @@ int main() {
 // CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK11-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
-// CHECK11-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK11-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK11-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK11-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81() #[[ATTR2]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK11-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK11:       omp_offload.failed2:
+// CHECK11-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK11-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK11-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK11-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK11-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK11:       omp_offload.failed3:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90() #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK11:       omp_offload.cont3:
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK11:       omp_offload.cont4:
+// CHECK11-NEXT:    [[TMP20:%.*]] = load i32, i32* @Arg, align 4
 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK11-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK11-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK11-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK11-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK11-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK11-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK11:       omp_offload.failed5:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK11:       omp_offload.cont6:
-// CHECK11-NEXT:    [[TMP15:%.*]] = load i32, i32* @Arg, align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP15]])
+// CHECK11-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK11-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK11-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK11-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK11:       omp_offload.failed7:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK11:       omp_offload.cont8:
+// CHECK11-NEXT:    [[TMP39:%.*]] = load i32, i32* @Arg, align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP39]])
 // CHECK11-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -6276,46 +6769,97 @@ int main() {
 // CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK11-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
-// CHECK11-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l59.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK11-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK11-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l59.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK11-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l59() #[[ATTR2]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l65.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK11-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK11:       omp_offload.failed2:
+// CHECK11-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK11-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK11-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l65.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK11-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK11-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK11:       omp_offload.failed3:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l65() #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK11:       omp_offload.cont3:
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK11:       omp_offload.cont4:
+// CHECK11-NEXT:    [[TMP20:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK11-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK11-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK11-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK11-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK11-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK11-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK11:       omp_offload.failed5:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK11:       omp_offload.cont6:
+// CHECK11-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK11-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK11-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK11-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK11:       omp_offload.failed7:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l71(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK11:       omp_offload.cont8:
 // CHECK11-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/distribute_parallel_for_simd_lastprivate_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_simd_lastprivate_codegen.cpp
index 457e83b..456e632 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_simd_lastprivate_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_simd_lastprivate_codegen.cpp
@@ -899,9 +899,26 @@ int main() {
 // CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l125.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK9-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l125.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK9-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l125(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -909,18 +926,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done3:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP39]]
+// CHECK9-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1377,27 +1394,44 @@ int main() {
 // CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK9-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK9-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP32]]
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1930,9 +1964,26 @@ int main() {
 // CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l125.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK11-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l125.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK11-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l125(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1940,18 +1991,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP39]]
+// CHECK11-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -2399,27 +2450,44 @@ int main() {
 // CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK11-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK11-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP32]]
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/distribute_parallel_for_simd_num_threads_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_simd_num_threads_codegen.cpp
index 4bfd973..0f9763d 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_simd_num_threads_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_simd_num_threads_codegen.cpp
@@ -114,64 +114,98 @@ int main() {
 // CHECK1:       invoke.cont:
 // CHECK1-NEXT:    store i8 [[CALL]], i8* [[A]], align 1
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68() #[[ATTR6:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       lpad:
-// CHECK1-NEXT:    [[TMP2:%.*]] = landingpad { i8*, i32 }
+// CHECK1-NEXT:    [[TMP10:%.*]] = landingpad { i8*, i32 }
 // CHECK1-NEXT:    cleanup
-// CHECK1-NEXT:    [[TMP3:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 0
-// CHECK1-NEXT:    store i8* [[TMP3]], i8** [[EXN_SLOT]], align 8
-// CHECK1-NEXT:    [[TMP4:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 1
-// CHECK1-NEXT:    store i32 [[TMP4]], i32* [[EHSELECTOR_SLOT]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 0
+// CHECK1-NEXT:    store i8* [[TMP11]], i8** [[EXN_SLOT]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 1
+// CHECK1-NEXT:    store i32 [[TMP12]], i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK1-NEXT:    call void @_ZN1SD1Ev(%struct.S* nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
 // CHECK1-NEXT:    br label [[EH_RESUME:%.*]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i8, i8* [[A]], align 1
+// CHECK1-NEXT:    [[TMP13:%.*]] = load i8, i8* [[A]], align 1
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i8*
-// CHECK1-NEXT:    store i8 [[TMP5]], i8* [[CONV]], align 1
-// CHECK1-NEXT:    [[TMP6:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP8]], align 8
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP11]], align 8
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    store i8 [[TMP13]], i8* [[CONV]], align 1
+// CHECK1-NEXT:    [[TMP14:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, i32 1, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK1-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP6]]) #[[ATTR6]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
-// CHECK1-NEXT:    [[TMP16:%.*]] = load i8, i8* [[A]], align 1
-// CHECK1-NEXT:    [[CONV4:%.*]] = sext i8 [[TMP16]] to i32
-// CHECK1-NEXT:    [[CALL6:%.*]] = invoke i32 @_Z5tmainIcLi5EEiv()
-// CHECK1-NEXT:    to label [[INVOKE_CONT5:%.*]] unwind label [[LPAD]]
-// CHECK1:       invoke.cont5:
-// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV4]], [[CALL6]]
-// CHECK1-NEXT:    [[CALL8:%.*]] = invoke i32 @_Z5tmainI1SLi1EEiv()
-// CHECK1-NEXT:    to label [[INVOKE_CONT7:%.*]] unwind label [[LPAD]]
-// CHECK1:       invoke.cont7:
-// CHECK1-NEXT:    [[ADD9:%.*]] = add nsw i32 [[ADD]], [[CALL8]]
-// CHECK1-NEXT:    store i32 [[ADD9]], i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK1-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP14]]) #[[ATTR6]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
+// CHECK1-NEXT:    [[TMP32:%.*]] = load i8, i8* [[A]], align 1
+// CHECK1-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP32]] to i32
+// CHECK1-NEXT:    [[CALL7:%.*]] = invoke i32 @_Z5tmainIcLi5EEiv()
+// CHECK1-NEXT:    to label [[INVOKE_CONT6:%.*]] unwind label [[LPAD]]
+// CHECK1:       invoke.cont6:
+// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV5]], [[CALL7]]
+// CHECK1-NEXT:    [[CALL9:%.*]] = invoke i32 @_Z5tmainI1SLi1EEiv()
+// CHECK1-NEXT:    to label [[INVOKE_CONT8:%.*]] unwind label [[LPAD]]
+// CHECK1:       invoke.cont8:
+// CHECK1-NEXT:    [[ADD10:%.*]] = add nsw i32 [[ADD]], [[CALL9]]
+// CHECK1-NEXT:    store i32 [[ADD10]], i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @_ZN1SD1Ev(%struct.S* nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
-// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP17]]
+// CHECK1-NEXT:    [[TMP33:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP33]]
 // CHECK1:       eh.resume:
 // CHECK1-NEXT:    [[EXN:%.*]] = load i8*, i8** [[EXN_SLOT]], align 8
 // CHECK1-NEXT:    [[SEL:%.*]] = load i32, i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK1-NEXT:    [[LPAD_VAL:%.*]] = insertvalue { i8*, i32 } undef, i8* [[EXN]], 0
-// CHECK1-NEXT:    [[LPAD_VAL10:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
-// CHECK1-NEXT:    resume { i8*, i32 } [[LPAD_VAL10]]
+// CHECK1-NEXT:    [[LPAD_VAL11:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
+// CHECK1-NEXT:    resume { i8*, i32 } [[LPAD_VAL11]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN1SC1El
@@ -543,21 +577,55 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52() #[[ATTR6]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57() #[[ATTR6]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret i32 0
 //
 //
@@ -567,21 +635,55 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52() #[[ATTR6]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57() #[[ATTR6]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret i32 0
 //
 //
@@ -1644,64 +1746,98 @@ int main() {
 // CHECK5:       invoke.cont:
 // CHECK5-NEXT:    store i8 [[CALL]], i8* [[A]], align 1
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK5-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK5-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK5-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK5-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK5-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK5-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68() #[[ATTR6:[0-9]+]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       lpad:
-// CHECK5-NEXT:    [[TMP2:%.*]] = landingpad { i8*, i32 }
+// CHECK5-NEXT:    [[TMP10:%.*]] = landingpad { i8*, i32 }
 // CHECK5-NEXT:    cleanup
-// CHECK5-NEXT:    [[TMP3:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 0
-// CHECK5-NEXT:    store i8* [[TMP3]], i8** [[EXN_SLOT]], align 8
-// CHECK5-NEXT:    [[TMP4:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 1
-// CHECK5-NEXT:    store i32 [[TMP4]], i32* [[EHSELECTOR_SLOT]], align 4
+// CHECK5-NEXT:    [[TMP11:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 0
+// CHECK5-NEXT:    store i8* [[TMP11]], i8** [[EXN_SLOT]], align 8
+// CHECK5-NEXT:    [[TMP12:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 1
+// CHECK5-NEXT:    store i32 [[TMP12]], i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK5-NEXT:    call void @_ZN1SD1Ev(%struct.S* nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
 // CHECK5-NEXT:    br label [[EH_RESUME:%.*]]
 // CHECK5:       omp_offload.cont:
-// CHECK5-NEXT:    [[TMP5:%.*]] = load i8, i8* [[A]], align 1
+// CHECK5-NEXT:    [[TMP13:%.*]] = load i8, i8* [[A]], align 1
 // CHECK5-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i8*
-// CHECK5-NEXT:    store i8 [[TMP5]], i8* [[CONV]], align 1
-// CHECK5-NEXT:    [[TMP6:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP6]], i64* [[TMP8]], align 8
-// CHECK5-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP6]], i64* [[TMP10]], align 8
-// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP11]], align 8
-// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK5-NEXT:    store i8 [[TMP13]], i8* [[CONV]], align 1
+// CHECK5-NEXT:    [[TMP14:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP14]], i64* [[TMP16]], align 8
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP14]], i64* [[TMP18]], align 8
+// CHECK5-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP19]], align 8
+// CHECK5-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, i32 1, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK5-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK5:       omp_offload.failed2:
-// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP6]]) #[[ATTR6]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK5:       omp_offload.cont3:
-// CHECK5-NEXT:    [[TMP16:%.*]] = load i8, i8* [[A]], align 1
-// CHECK5-NEXT:    [[CONV4:%.*]] = sext i8 [[TMP16]] to i32
-// CHECK5-NEXT:    [[CALL6:%.*]] = invoke i32 @_Z5tmainIcLi5EEiv()
-// CHECK5-NEXT:    to label [[INVOKE_CONT5:%.*]] unwind label [[LPAD]]
-// CHECK5:       invoke.cont5:
-// CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV4]], [[CALL6]]
-// CHECK5-NEXT:    [[CALL8:%.*]] = invoke i32 @_Z5tmainI1SLi1EEiv()
-// CHECK5-NEXT:    to label [[INVOKE_CONT7:%.*]] unwind label [[LPAD]]
-// CHECK5:       invoke.cont7:
-// CHECK5-NEXT:    [[ADD9:%.*]] = add nsw i32 [[ADD]], [[CALL8]]
-// CHECK5-NEXT:    store i32 [[ADD9]], i32* [[RETVAL]], align 4
+// CHECK5-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK5-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK5-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK5-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK5:       omp_offload.failed3:
+// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP14]]) #[[ATTR6]]
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK5:       omp_offload.cont4:
+// CHECK5-NEXT:    [[TMP32:%.*]] = load i8, i8* [[A]], align 1
+// CHECK5-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP32]] to i32
+// CHECK5-NEXT:    [[CALL7:%.*]] = invoke i32 @_Z5tmainIcLi5EEiv()
+// CHECK5-NEXT:    to label [[INVOKE_CONT6:%.*]] unwind label [[LPAD]]
+// CHECK5:       invoke.cont6:
+// CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV5]], [[CALL7]]
+// CHECK5-NEXT:    [[CALL9:%.*]] = invoke i32 @_Z5tmainI1SLi1EEiv()
+// CHECK5-NEXT:    to label [[INVOKE_CONT8:%.*]] unwind label [[LPAD]]
+// CHECK5:       invoke.cont8:
+// CHECK5-NEXT:    [[ADD10:%.*]] = add nsw i32 [[ADD]], [[CALL9]]
+// CHECK5-NEXT:    store i32 [[ADD10]], i32* [[RETVAL]], align 4
 // CHECK5-NEXT:    call void @_ZN1SD1Ev(%struct.S* nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
-// CHECK5-NEXT:    [[TMP17:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP17]]
+// CHECK5-NEXT:    [[TMP33:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP33]]
 // CHECK5:       eh.resume:
 // CHECK5-NEXT:    [[EXN:%.*]] = load i8*, i8** [[EXN_SLOT]], align 8
 // CHECK5-NEXT:    [[SEL:%.*]] = load i32, i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK5-NEXT:    [[LPAD_VAL:%.*]] = insertvalue { i8*, i32 } undef, i8* [[EXN]], 0
-// CHECK5-NEXT:    [[LPAD_VAL10:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
-// CHECK5-NEXT:    resume { i8*, i32 } [[LPAD_VAL10]]
+// CHECK5-NEXT:    [[LPAD_VAL11:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
+// CHECK5-NEXT:    resume { i8*, i32 } [[LPAD_VAL11]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@_ZN1SC1El
@@ -2073,21 +2209,55 @@ int main() {
 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK5-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK5-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK5-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK5-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK5-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK5-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52() #[[ATTR6]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK5-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK5:       omp_offload.failed2:
+// CHECK5-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK5-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK5-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK5-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK5-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK5:       omp_offload.failed3:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57() #[[ATTR6]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK5:       omp_offload.cont3:
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK5:       omp_offload.cont4:
 // CHECK5-NEXT:    ret i32 0
 //
 //
@@ -2097,21 +2267,55 @@ int main() {
 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK5-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK5-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK5-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK5-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK5-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK5-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52() #[[ATTR6]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK5-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK5:       omp_offload.failed2:
+// CHECK5-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK5-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK5-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK5-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK5-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK5:       omp_offload.failed3:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57() #[[ATTR6]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK5:       omp_offload.cont3:
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK5:       omp_offload.cont4:
 // CHECK5-NEXT:    ret i32 0
 //
 //
@@ -2835,64 +3039,98 @@ int main() {
 // CHECK9:       invoke.cont:
 // CHECK9-NEXT:    store i8 [[CALL]], i8* [[A]], align 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68() #[[ATTR6:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       lpad:
-// CHECK9-NEXT:    [[TMP2:%.*]] = landingpad { i8*, i32 }
+// CHECK9-NEXT:    [[TMP10:%.*]] = landingpad { i8*, i32 }
 // CHECK9-NEXT:    cleanup
-// CHECK9-NEXT:    [[TMP3:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 0
-// CHECK9-NEXT:    store i8* [[TMP3]], i8** [[EXN_SLOT]], align 8
-// CHECK9-NEXT:    [[TMP4:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 1
-// CHECK9-NEXT:    store i32 [[TMP4]], i32* [[EHSELECTOR_SLOT]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 0
+// CHECK9-NEXT:    store i8* [[TMP11]], i8** [[EXN_SLOT]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 1
+// CHECK9-NEXT:    store i32 [[TMP12]], i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK9-NEXT:    call void @_ZN1SD1Ev(%struct.S* nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
 // CHECK9-NEXT:    br label [[EH_RESUME:%.*]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP5:%.*]] = load i8, i8* [[A]], align 1
+// CHECK9-NEXT:    [[TMP13:%.*]] = load i8, i8* [[A]], align 1
 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i8*
-// CHECK9-NEXT:    store i8 [[TMP5]], i8* [[CONV]], align 1
-// CHECK9-NEXT:    [[TMP6:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP6]], i64* [[TMP8]], align 8
-// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP6]], i64* [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP11]], align 8
-// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    store i8 [[TMP13]], i8* [[CONV]], align 1
+// CHECK9-NEXT:    [[TMP14:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP14]], i64* [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP14]], i64* [[TMP18]], align 8
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP19]], align 8
+// CHECK9-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, i32 1, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK9-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK9:       omp_offload.failed2:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP6]]) #[[ATTR6]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK9:       omp_offload.cont3:
-// CHECK9-NEXT:    [[TMP16:%.*]] = load i8, i8* [[A]], align 1
-// CHECK9-NEXT:    [[CONV4:%.*]] = sext i8 [[TMP16]] to i32
-// CHECK9-NEXT:    [[CALL6:%.*]] = invoke i32 @_Z5tmainIcLi5EEiv()
-// CHECK9-NEXT:    to label [[INVOKE_CONT5:%.*]] unwind label [[LPAD]]
-// CHECK9:       invoke.cont5:
-// CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV4]], [[CALL6]]
-// CHECK9-NEXT:    [[CALL8:%.*]] = invoke i32 @_Z5tmainI1SLi1EEiv()
-// CHECK9-NEXT:    to label [[INVOKE_CONT7:%.*]] unwind label [[LPAD]]
-// CHECK9:       invoke.cont7:
-// CHECK9-NEXT:    [[ADD9:%.*]] = add nsw i32 [[ADD]], [[CALL8]]
-// CHECK9-NEXT:    store i32 [[ADD9]], i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK9:       omp_offload.failed3:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP14]]) #[[ATTR6]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK9:       omp_offload.cont4:
+// CHECK9-NEXT:    [[TMP32:%.*]] = load i8, i8* [[A]], align 1
+// CHECK9-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP32]] to i32
+// CHECK9-NEXT:    [[CALL7:%.*]] = invoke i32 @_Z5tmainIcLi5EEiv()
+// CHECK9-NEXT:    to label [[INVOKE_CONT6:%.*]] unwind label [[LPAD]]
+// CHECK9:       invoke.cont6:
+// CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV5]], [[CALL7]]
+// CHECK9-NEXT:    [[CALL9:%.*]] = invoke i32 @_Z5tmainI1SLi1EEiv()
+// CHECK9-NEXT:    to label [[INVOKE_CONT8:%.*]] unwind label [[LPAD]]
+// CHECK9:       invoke.cont8:
+// CHECK9-NEXT:    [[ADD10:%.*]] = add nsw i32 [[ADD]], [[CALL9]]
+// CHECK9-NEXT:    store i32 [[ADD10]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    call void @_ZN1SD1Ev(%struct.S* nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
-// CHECK9-NEXT:    [[TMP17:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP17]]
+// CHECK9-NEXT:    [[TMP33:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP33]]
 // CHECK9:       eh.resume:
 // CHECK9-NEXT:    [[EXN:%.*]] = load i8*, i8** [[EXN_SLOT]], align 8
 // CHECK9-NEXT:    [[SEL:%.*]] = load i32, i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK9-NEXT:    [[LPAD_VAL:%.*]] = insertvalue { i8*, i32 } undef, i8* [[EXN]], 0
-// CHECK9-NEXT:    [[LPAD_VAL10:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
-// CHECK9-NEXT:    resume { i8*, i32 } [[LPAD_VAL10]]
+// CHECK9-NEXT:    [[LPAD_VAL11:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
+// CHECK9-NEXT:    resume { i8*, i32 } [[LPAD_VAL11]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SC1El
@@ -3264,21 +3502,55 @@ int main() {
 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52() #[[ATTR6]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK9-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK9:       omp_offload.failed2:
+// CHECK9-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK9-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK9-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK9:       omp_offload.failed3:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57() #[[ATTR6]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK9:       omp_offload.cont3:
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK9:       omp_offload.cont4:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -3288,21 +3560,55 @@ int main() {
 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52() #[[ATTR6]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK9-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK9:       omp_offload.failed2:
+// CHECK9-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK9-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK9-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK9:       omp_offload.failed3:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57() #[[ATTR6]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK9:       omp_offload.cont3:
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK9:       omp_offload.cont4:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -4365,64 +4671,98 @@ int main() {
 // CHECK13:       invoke.cont:
 // CHECK13-NEXT:    store i8 [[CALL]], i8* [[A]], align 1
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK13-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK13-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK13-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK13-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK13-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK13-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK13-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK13-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK13-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK13-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK13-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK13-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68() #[[ATTR6:[0-9]+]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       lpad:
-// CHECK13-NEXT:    [[TMP2:%.*]] = landingpad { i8*, i32 }
+// CHECK13-NEXT:    [[TMP10:%.*]] = landingpad { i8*, i32 }
 // CHECK13-NEXT:    cleanup
-// CHECK13-NEXT:    [[TMP3:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 0
-// CHECK13-NEXT:    store i8* [[TMP3]], i8** [[EXN_SLOT]], align 8
-// CHECK13-NEXT:    [[TMP4:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 1
-// CHECK13-NEXT:    store i32 [[TMP4]], i32* [[EHSELECTOR_SLOT]], align 4
+// CHECK13-NEXT:    [[TMP11:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 0
+// CHECK13-NEXT:    store i8* [[TMP11]], i8** [[EXN_SLOT]], align 8
+// CHECK13-NEXT:    [[TMP12:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 1
+// CHECK13-NEXT:    store i32 [[TMP12]], i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK13-NEXT:    call void @_ZN1SD1Ev(%struct.S* nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
 // CHECK13-NEXT:    br label [[EH_RESUME:%.*]]
 // CHECK13:       omp_offload.cont:
-// CHECK13-NEXT:    [[TMP5:%.*]] = load i8, i8* [[A]], align 1
+// CHECK13-NEXT:    [[TMP13:%.*]] = load i8, i8* [[A]], align 1
 // CHECK13-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i8*
-// CHECK13-NEXT:    store i8 [[TMP5]], i8* [[CONV]], align 1
-// CHECK13-NEXT:    [[TMP6:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK13-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP6]], i64* [[TMP8]], align 8
-// CHECK13-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP6]], i64* [[TMP10]], align 8
-// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP11]], align 8
-// CHECK13-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK13-NEXT:    store i8 [[TMP13]], i8* [[CONV]], align 1
+// CHECK13-NEXT:    [[TMP14:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK13-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP14]], i64* [[TMP16]], align 8
+// CHECK13-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP14]], i64* [[TMP18]], align 8
+// CHECK13-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP19]], align 8
+// CHECK13-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK13-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, i32 1, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK13-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK13-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK13:       omp_offload.failed2:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP6]]) #[[ATTR6]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK13:       omp_offload.cont3:
-// CHECK13-NEXT:    [[TMP16:%.*]] = load i8, i8* [[A]], align 1
-// CHECK13-NEXT:    [[CONV4:%.*]] = sext i8 [[TMP16]] to i32
-// CHECK13-NEXT:    [[CALL6:%.*]] = invoke i32 @_Z5tmainIcLi5EEiv()
-// CHECK13-NEXT:    to label [[INVOKE_CONT5:%.*]] unwind label [[LPAD]]
-// CHECK13:       invoke.cont5:
-// CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV4]], [[CALL6]]
-// CHECK13-NEXT:    [[CALL8:%.*]] = invoke i32 @_Z5tmainI1SLi1EEiv()
-// CHECK13-NEXT:    to label [[INVOKE_CONT7:%.*]] unwind label [[LPAD]]
-// CHECK13:       invoke.cont7:
-// CHECK13-NEXT:    [[ADD9:%.*]] = add nsw i32 [[ADD]], [[CALL8]]
-// CHECK13-NEXT:    store i32 [[ADD9]], i32* [[RETVAL]], align 4
+// CHECK13-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK13-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK13-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK13-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK13-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK13-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK13-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK13-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK13-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK13-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK13-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK13:       omp_offload.failed3:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l74(i64 [[TMP14]]) #[[ATTR6]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK13:       omp_offload.cont4:
+// CHECK13-NEXT:    [[TMP32:%.*]] = load i8, i8* [[A]], align 1
+// CHECK13-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP32]] to i32
+// CHECK13-NEXT:    [[CALL7:%.*]] = invoke i32 @_Z5tmainIcLi5EEiv()
+// CHECK13-NEXT:    to label [[INVOKE_CONT6:%.*]] unwind label [[LPAD]]
+// CHECK13:       invoke.cont6:
+// CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV5]], [[CALL7]]
+// CHECK13-NEXT:    [[CALL9:%.*]] = invoke i32 @_Z5tmainI1SLi1EEiv()
+// CHECK13-NEXT:    to label [[INVOKE_CONT8:%.*]] unwind label [[LPAD]]
+// CHECK13:       invoke.cont8:
+// CHECK13-NEXT:    [[ADD10:%.*]] = add nsw i32 [[ADD]], [[CALL9]]
+// CHECK13-NEXT:    store i32 [[ADD10]], i32* [[RETVAL]], align 4
 // CHECK13-NEXT:    call void @_ZN1SD1Ev(%struct.S* nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
-// CHECK13-NEXT:    [[TMP17:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK13-NEXT:    ret i32 [[TMP17]]
+// CHECK13-NEXT:    [[TMP33:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK13-NEXT:    ret i32 [[TMP33]]
 // CHECK13:       eh.resume:
 // CHECK13-NEXT:    [[EXN:%.*]] = load i8*, i8** [[EXN_SLOT]], align 8
 // CHECK13-NEXT:    [[SEL:%.*]] = load i32, i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK13-NEXT:    [[LPAD_VAL:%.*]] = insertvalue { i8*, i32 } undef, i8* [[EXN]], 0
-// CHECK13-NEXT:    [[LPAD_VAL10:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
-// CHECK13-NEXT:    resume { i8*, i32 } [[LPAD_VAL10]]
+// CHECK13-NEXT:    [[LPAD_VAL11:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
+// CHECK13-NEXT:    resume { i8*, i32 } [[LPAD_VAL11]]
 //
 //
 // CHECK13-LABEL: define {{[^@]+}}@_ZN1SC1El
@@ -4794,21 +5134,55 @@ int main() {
 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK13-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK13-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK13-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK13-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK13-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK13-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK13-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK13-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK13-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK13-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK13-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK13-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l52() #[[ATTR6]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       omp_offload.cont:
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK13-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK13-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK13-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK13:       omp_offload.failed2:
+// CHECK13-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK13-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK13-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK13-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK13-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK13-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK13-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK13-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK13-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK13:       omp_offload.failed3:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l57() #[[ATTR6]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK13:       omp_offload.cont3:
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK13:       omp_offload.cont4:
 // CHECK13-NEXT:    ret i32 0
 //
 //
@@ -4818,21 +5192,55 @@ int main() {
 // CHECK13-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK13-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK13-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK13-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK13-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK13-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK13-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK13-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK13-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK13-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK13-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK13-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK13-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l52() #[[ATTR6]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       omp_offload.cont:
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK13-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK13-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK13-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK13:       omp_offload.failed2:
+// CHECK13-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK13-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK13-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK13-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK13-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK13-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK13-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK13-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK13-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK13:       omp_offload.failed3:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l57() #[[ATTR6]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK13:       omp_offload.cont3:
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK13:       omp_offload.cont4:
 // CHECK13-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/distribute_parallel_for_simd_private_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_simd_private_codegen.cpp
index b69a5e2..3a8b78e 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_simd_private_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_simd_private_codegen.cpp
@@ -589,9 +589,26 @@ int main() {
 // CHECK9-NEXT:    store %struct.S* [[TEST]], %struct.S** [[VAR]], align 8
 // CHECK9-NEXT:    store %struct.S* undef, %struct.S** [[_TMP1]], align 8
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l95.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK9-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK9-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l95.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK9-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l95() #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -599,18 +616,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP4]]
+// CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -901,27 +918,44 @@ int main() {
 // CHECK9-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 8
 // CHECK9-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 8
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK9-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK9-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK9-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49() #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP4]]
+// CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1289,9 +1323,26 @@ int main() {
 // CHECK11-NEXT:    store %struct.S* [[TEST]], %struct.S** [[VAR]], align 4
 // CHECK11-NEXT:    store %struct.S* undef, %struct.S** [[_TMP1]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l95.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK11-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l95.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK11-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l95() #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1299,18 +1350,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP4]]
+// CHECK11-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1595,27 +1646,44 @@ int main() {
 // CHECK11-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 4
 // CHECK11-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK11-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK11-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49() #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP4]]
+// CHECK11-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/distribute_parallel_for_simd_proc_bind_codegen.cpp b/clang/test/OpenMP/distribute_parallel_for_simd_proc_bind_codegen.cpp
index 16000c7..8df0cc3 100644 (file)
--- a/clang/test/OpenMP/distribute_parallel_for_simd_proc_bind_codegen.cpp
+++ b/clang/test/OpenMP/distribute_parallel_for_simd_proc_bind_codegen.cpp
@@ -61,21 +61,55 @@ int main() {
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l37.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l37.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l37() #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l41.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l41.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l41() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK1-NEXT:    ret i32 [[CALL]]
 //
@@ -389,9 +423,26 @@ int main() {
 // CHECK1-NEXT:  entry:
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l29.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l29.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l29() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/distribute_private_codegen.cpp b/clang/test/OpenMP/distribute_private_codegen.cpp
index 9ddd5ac..c242062 100644 (file)
--- a/clang/test/OpenMP/distribute_private_codegen.cpp
+++ b/clang/test/OpenMP/distribute_private_codegen.cpp
@@ -364,36 +364,70 @@ int main() {
 // CHECK9-NEXT:    store %struct.S* [[TEST]], %struct.S** [[VAR]], align 8
 // CHECK9-NEXT:    store %struct.S* undef, %struct.S** [[_TMP1]], align 8
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK9-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK9-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK9-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93() #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP3:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP4:%.*]] = icmp ne i32 [[TMP3]], 0
-// CHECK9-NEXT:    br i1 [[TMP4]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
-// CHECK9:       omp_offload.failed3:
+// CHECK9-NEXT:    [[KERNEL_ARGS3:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP12]], align 4
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP18]], align 8
+// CHECK9-NEXT:    [[TMP19:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]])
+// CHECK9-NEXT:    [[TMP20:%.*]] = icmp ne i32 [[TMP19]], 0
+// CHECK9-NEXT:    br i1 [[TMP20]], label [[OMP_OFFLOAD_FAILED4:%.*]], label [[OMP_OFFLOAD_CONT5:%.*]]
+// CHECK9:       omp_offload.failed4:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102() #[[ATTR4]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
-// CHECK9:       omp_offload.cont4:
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT5]]
+// CHECK9:       omp_offload.cont5:
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP5]], [[OMP_OFFLOAD_CONT4]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP21]], [[OMP_OFFLOAD_CONT5]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
-// CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE5:%.*]], label [[ARRAYDESTROY_BODY]]
-// CHECK9:       arraydestroy.done5:
+// CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE6:%.*]], label [[ARRAYDESTROY_BODY]]
+// CHECK9:       arraydestroy.done6:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP6:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP6]]
+// CHECK9-NEXT:    [[TMP22:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP22]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -633,27 +667,44 @@ int main() {
 // CHECK9-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 8
 // CHECK9-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 8
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK9-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK9-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK9-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49() #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP4]]
+// CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -905,36 +956,70 @@ int main() {
 // CHECK11-NEXT:    store %struct.S* [[TEST]], %struct.S** [[VAR]], align 4
 // CHECK11-NEXT:    store %struct.S* undef, %struct.S** [[_TMP1]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK11-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK11-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93() #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP3:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP4:%.*]] = icmp ne i32 [[TMP3]], 0
-// CHECK11-NEXT:    br i1 [[TMP4]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
-// CHECK11:       omp_offload.failed3:
+// CHECK11-NEXT:    [[KERNEL_ARGS3:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP15]], align 4
+// CHECK11-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP16]], align 4
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP17]], align 4
+// CHECK11-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP18]], align 4
+// CHECK11-NEXT:    [[TMP19:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]])
+// CHECK11-NEXT:    [[TMP20:%.*]] = icmp ne i32 [[TMP19]], 0
+// CHECK11-NEXT:    br i1 [[TMP20]], label [[OMP_OFFLOAD_FAILED4:%.*]], label [[OMP_OFFLOAD_CONT5:%.*]]
+// CHECK11:       omp_offload.failed4:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102() #[[ATTR4]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
-// CHECK11:       omp_offload.cont4:
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT5]]
+// CHECK11:       omp_offload.cont5:
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP5]], [[OMP_OFFLOAD_CONT4]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP21]], [[OMP_OFFLOAD_CONT5]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
-// CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE5:%.*]], label [[ARRAYDESTROY_BODY]]
-// CHECK11:       arraydestroy.done5:
+// CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE6:%.*]], label [[ARRAYDESTROY_BODY]]
+// CHECK11:       arraydestroy.done6:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP6:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP6]]
+// CHECK11-NEXT:    [[TMP22:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP22]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1172,27 +1257,44 @@ int main() {
 // CHECK11-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 4
 // CHECK11-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK11-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK11-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49() #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP4]]
+// CHECK11-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/distribute_simd_codegen.cpp b/clang/test/OpenMP/distribute_simd_codegen.cpp
index 349d920..cbde85c 100644 (file)
--- a/clang/test/OpenMP/distribute_simd_codegen.cpp
+++ b/clang/test/OpenMP/distribute_simd_codegen.cpp
@@ -198,9 +198,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 4571424)
-// CHECK1-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l70.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK1-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l70.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK1-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l70(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR3:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -380,9 +397,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 4571424)
-// CHECK1-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l86.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK1-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l86.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK1-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l86(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -560,9 +594,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 16908289)
-// CHECK1-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l103.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK1-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l103.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK1-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l103(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -757,9 +808,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK1-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP18]], 1
 // CHECK1-NEXT:    [[TMP19:%.*]] = zext i32 [[ADD6]] to i64
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP19]])
-// CHECK1-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l115.region_id, i32 2, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
-// CHECK1-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP20]], align 4
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP21]], align 4
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP14]], i8*** [[TMP22]], align 8
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP15]], i8*** [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l115.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
+// CHECK1-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l115(i64 [[TMP1]], i64 [[TMP3]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -926,9 +994,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l135.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l135.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l135(i64 [[TMP1]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1096,9 +1181,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 4571424)
-// CHECK3-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l70.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK3-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l70.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK3-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l70(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR3:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1274,9 +1376,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 4571424)
-// CHECK3-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l86.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK3-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l86.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK3-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l86(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1450,9 +1569,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 16908289)
-// CHECK3-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l103.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK3-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l103.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK3-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l103(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1643,9 +1779,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK3-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP18]], 1
 // CHECK3-NEXT:    [[TMP19:%.*]] = zext i32 [[ADD6]] to i64
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP19]])
-// CHECK3-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l115.region_id, i32 2, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
-// CHECK3-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP14]], i8*** [[TMP22]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP15]], i8*** [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l115.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
+// CHECK3-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l115(i32 [[TMP1]], i32 [[TMP3]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1812,9 +1965,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l135.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l135.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK3-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l135(i32 [[TMP1]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1982,9 +2152,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 4571424)
-// CHECK5-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l70.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK5-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 8
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP30]], align 8
+// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP32]], align 8
+// CHECK5-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK5-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l70.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK5-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l70(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR3:[0-9]+]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2164,9 +2351,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 4571424)
-// CHECK5-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l86.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK5-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 8
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP30]], align 8
+// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP32]], align 8
+// CHECK5-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK5-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l86.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK5-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l86(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR3]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2344,9 +2548,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 16908289)
-// CHECK5-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l103.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK5-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 8
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP30]], align 8
+// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP32]], align 8
+// CHECK5-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK5-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l103.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK5-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l103(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR3]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2541,9 +2762,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK5-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP18]], 1
 // CHECK5-NEXT:    [[TMP19:%.*]] = zext i32 [[ADD6]] to i64
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP19]])
-// CHECK5-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l115.region_id, i32 2, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
-// CHECK5-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP20]], align 4
+// CHECK5-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 2, i32* [[TMP21]], align 4
+// CHECK5-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP14]], i8*** [[TMP22]], align 8
+// CHECK5-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP15]], i8*** [[TMP23]], align 8
+// CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP24]], align 8
+// CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP25]], align 8
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP26]], align 8
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP27]], align 8
+// CHECK5-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l115.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
+// CHECK5-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l115(i64 [[TMP1]], i64 [[TMP3]]) #[[ATTR3]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2741,9 +2979,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l135.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK5-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK5-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK5-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK5-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l135.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK5-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l135(i64 [[TMP1]]) #[[ATTR3]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2911,9 +3166,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK7-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 4571424)
-// CHECK7-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l70.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK7-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK7-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK7-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 4
+// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK7-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP30]], align 4
+// CHECK7-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK7-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP32]], align 4
+// CHECK7-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK7-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l70.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK7-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z23without_schedule_clausePfS_S_S__l70(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR3:[0-9]+]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3089,9 +3361,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK7-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 4571424)
-// CHECK7-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l86.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK7-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK7-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK7-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 4
+// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK7-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP30]], align 4
+// CHECK7-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK7-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP32]], align 4
+// CHECK7-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK7-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l86.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK7-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18static_not_chunkedPfS_S_S__l86(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR3]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3265,9 +3554,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK7-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 16908289)
-// CHECK7-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l103.region_id, i32 4, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK7-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK7-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 4, i32* [[TMP27]], align 4
+// CHECK7-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP24]], i8*** [[TMP28]], align 4
+// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK7-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP30]], align 4
+// CHECK7-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK7-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP32]], align 4
+// CHECK7-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK7-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l103.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK7-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14static_chunkedPfS_S_S__l103(float* [[TMP0]], float* [[TMP1]], float* [[TMP2]], float* [[TMP3]]) #[[ATTR3]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3458,9 +3764,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK7-NEXT:    [[ADD6:%.*]] = add nsw i32 [[TMP18]], 1
 // CHECK7-NEXT:    [[TMP19:%.*]] = zext i32 [[ADD6]] to i64
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP19]])
-// CHECK7-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l115.region_id, i32 2, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK7-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
-// CHECK7-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP20]], align 4
+// CHECK7-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 2, i32* [[TMP21]], align 4
+// CHECK7-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP14]], i8*** [[TMP22]], align 4
+// CHECK7-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP15]], i8*** [[TMP23]], align 4
+// CHECK7-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP24]], align 4
+// CHECK7-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP25]], align 4
+// CHECK7-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP26]], align 4
+// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP27]], align 4
+// CHECK7-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l115.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
+// CHECK7-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z12test_precondv_l115(i32 [[TMP1]], i32 [[TMP3]]) #[[ATTR3]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3658,9 +3981,26 @@ int fint(void) { return ftemplate<int>(); }
 // CHECK7-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 100)
-// CHECK7-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l135.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK7-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK7-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK7-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK7-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK7-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK7-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK7-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK7-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK7-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK7-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l135.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK7-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_v_l135(i32 [[TMP1]]) #[[ATTR3]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/distribute_simd_firstprivate_codegen.cpp b/clang/test/OpenMP/distribute_simd_firstprivate_codegen.cpp
index fc40097..3353b36 100644 (file)
--- a/clang/test/OpenMP/distribute_simd_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/distribute_simd_firstprivate_codegen.cpp
@@ -626,9 +626,26 @@ int main() {
 // CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK9-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK9-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -636,18 +653,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done3:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP39]]
+// CHECK9-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -924,27 +941,44 @@ int main() {
 // CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK9-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK9-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP32]]
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1302,9 +1336,26 @@ int main() {
 // CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK11-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK11-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1312,18 +1363,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP39]]
+// CHECK11-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1595,27 +1646,44 @@ int main() {
 // CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK11-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK11-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP32]]
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/distribute_simd_lastprivate_codegen.cpp b/clang/test/OpenMP/distribute_simd_lastprivate_codegen.cpp
index e8b1ecb..01ae013 100644 (file)
--- a/clang/test/OpenMP/distribute_simd_lastprivate_codegen.cpp
+++ b/clang/test/OpenMP/distribute_simd_lastprivate_codegen.cpp
@@ -613,9 +613,26 @@ int main() {
 // CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK9-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK9-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -623,18 +640,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done3:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP39]]
+// CHECK9-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -929,27 +946,44 @@ int main() {
 // CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK9-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK9-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP32]]
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1326,9 +1360,26 @@ int main() {
 // CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK11-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK11-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1336,18 +1387,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP39]]
+// CHECK11-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1637,27 +1688,44 @@ int main() {
 // CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK11-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK11-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP32]]
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/distribute_simd_private_codegen.cpp b/clang/test/OpenMP/distribute_simd_private_codegen.cpp
index 480ffd5..c12d14b 100644 (file)
--- a/clang/test/OpenMP/distribute_simd_private_codegen.cpp
+++ b/clang/test/OpenMP/distribute_simd_private_codegen.cpp
@@ -409,50 +409,84 @@ int main() {
 // CHECK9-NEXT:    store %struct.S* [[TEST]], %struct.S** [[VAR]], align 8
 // CHECK9-NEXT:    store %struct.S* undef, %struct.S** [[_TMP1]], align 8
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK9-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK9-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK9-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93() #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = load i32, i32* [[I]], align 4
 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[I_CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP3]], i32* [[CONV]], align 4
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i64, i64* [[I_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP4]], i64* [[TMP6]], align 8
-// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP4]], i64* [[TMP8]], align 8
-// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP9]], align 8
-// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 [[TMP11]], i32* [[CONV]], align 4
+// CHECK9-NEXT:    [[TMP12:%.*]] = load i64, i64* [[I_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP14:%.*]] = bitcast i8** [[TMP13]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP12]], i64* [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP12]], i64* [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
-// CHECK9-NEXT:    br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
-// CHECK9:       omp_offload.failed3:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i64 [[TMP4]]) #[[ATTR4]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
-// CHECK9:       omp_offload.cont4:
+// CHECK9-NEXT:    [[KERNEL_ARGS3:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP20]], align 4
+// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP21]], align 4
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP18]], i8*** [[TMP22]], align 8
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP19]], i8*** [[TMP23]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP24]], align 8
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP25]], align 8
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]])
+// CHECK9-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
+// CHECK9-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED4:%.*]], label [[OMP_OFFLOAD_CONT5:%.*]]
+// CHECK9:       omp_offload.failed4:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i64 [[TMP12]]) #[[ATTR4]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT5]]
+// CHECK9:       omp_offload.cont5:
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP14]], [[OMP_OFFLOAD_CONT4]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP30]], [[OMP_OFFLOAD_CONT5]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
-// CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE5:%.*]], label [[ARRAYDESTROY_BODY]]
-// CHECK9:       arraydestroy.done5:
+// CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE6:%.*]], label [[ARRAYDESTROY_BODY]]
+// CHECK9:       arraydestroy.done6:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP15:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP15]]
+// CHECK9-NEXT:    [[TMP31:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP31]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -711,27 +745,44 @@ int main() {
 // CHECK9-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 8
 // CHECK9-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 8
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK9-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK9-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK9-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49() #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP4]]
+// CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -994,49 +1045,83 @@ int main() {
 // CHECK11-NEXT:    store %struct.S* [[TEST]], %struct.S** [[VAR]], align 4
 // CHECK11-NEXT:    store %struct.S* undef, %struct.S** [[_TMP1]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK11-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK11-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93() #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP3:%.*]] = load i32, i32* [[I]], align 4
-// CHECK11-NEXT:    store i32 [[TMP3]], i32* [[I_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[I_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP4]], i32* [[TMP6]], align 4
-// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP4]], i32* [[TMP8]], align 4
-// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP9]], align 4
-// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP11:%.*]] = load i32, i32* [[I]], align 4
+// CHECK11-NEXT:    store i32 [[TMP11]], i32* [[I_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = load i32, i32* [[I_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP14:%.*]] = bitcast i8** [[TMP13]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP12]], i32* [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP12]], i32* [[TMP16]], align 4
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP17]], align 4
+// CHECK11-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
-// CHECK11-NEXT:    br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
-// CHECK11:       omp_offload.failed3:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i32 [[TMP4]]) #[[ATTR4]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
-// CHECK11:       omp_offload.cont4:
+// CHECK11-NEXT:    [[KERNEL_ARGS3:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP20]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP21]], align 4
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP18]], i8*** [[TMP22]], align 4
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP19]], i8*** [[TMP23]], align 4
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP24]], align 4
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP25]], align 4
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP26]], align 4
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP27]], align 4
+// CHECK11-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]])
+// CHECK11-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
+// CHECK11-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED4:%.*]], label [[OMP_OFFLOAD_CONT5:%.*]]
+// CHECK11:       omp_offload.failed4:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i32 [[TMP12]]) #[[ATTR4]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT5]]
+// CHECK11:       omp_offload.cont5:
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP14]], [[OMP_OFFLOAD_CONT4]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP30]], [[OMP_OFFLOAD_CONT5]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
-// CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE5:%.*]], label [[ARRAYDESTROY_BODY]]
-// CHECK11:       arraydestroy.done5:
+// CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE6:%.*]], label [[ARRAYDESTROY_BODY]]
+// CHECK11:       arraydestroy.done6:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP15:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP15]]
+// CHECK11-NEXT:    [[TMP31:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP31]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1292,27 +1377,44 @@ int main() {
 // CHECK11-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 4
 // CHECK11-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK11-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK11-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49() #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP4]]
+// CHECK11-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/distribute_simd_reduction_codegen.cpp b/clang/test/OpenMP/distribute_simd_reduction_codegen.cpp
index a72bde2..1e76274 100644 (file)
--- a/clang/test/OpenMP/distribute_simd_reduction_codegen.cpp
+++ b/clang/test/OpenMP/distribute_simd_reduction_codegen.cpp
@@ -107,9 +107,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l64.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l64.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l64(i64 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -236,9 +253,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP10:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[TMP10]], 0
-// CHECK1-NEXT:    br i1 [[TMP11]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP9]], i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i64 [[TMP2]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -367,9 +401,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l64.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l64.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK3-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l64(i32 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -494,9 +545,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP10:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[TMP10]], 0
-// CHECK3-NEXT:    br i1 [[TMP11]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP9]], i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK3-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i32 [[TMP2]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/nvptx_lambda_capturing.cpp b/clang/test/OpenMP/nvptx_lambda_capturing.cpp
index 4a12393..046f683 100644 (file)
--- a/clang/test/OpenMP/nvptx_lambda_capturing.cpp
+++ b/clang/test/OpenMP/nvptx_lambda_capturing.cpp
@@ -211,138 +211,172 @@ int main(int argc, char **argv) {
 // CHECK1-NEXT:    store i8* null, i8** [[TMP85]], align 8
 // CHECK1-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP88:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l41.region_id, i32 11, i8** [[TMP86]], i8** [[TMP87]], i64* getelementptr inbounds ([11 x i64], [11 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([11 x i64], [11 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null)
-// CHECK1-NEXT:    [[TMP89:%.*]] = icmp ne i32 [[TMP88]], 0
-// CHECK1-NEXT:    br i1 [[TMP89]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP88]], align 4
+// CHECK1-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 11, i32* [[TMP89]], align 4
+// CHECK1-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP86]], i8*** [[TMP90]], align 8
+// CHECK1-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP87]], i8*** [[TMP91]], align 8
+// CHECK1-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([11 x i64], [11 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP92]], align 8
+// CHECK1-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([11 x i64], [11 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP93]], align 8
+// CHECK1-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP94]], align 8
+// CHECK1-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP95]], align 8
+// CHECK1-NEXT:    [[TMP96:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l41.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP97:%.*]] = icmp ne i32 [[TMP96]], 0
+// CHECK1-NEXT:    br i1 [[TMP97]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l41(i64 [[TMP11]], i32* [[TMP12]], i32* [[TMP13]], i32* [[TMP14]], i32* [[A]], %class.anon* [[TMP15]]) #[[ATTR4:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP90:%.*]] = load i32*, i32** [[B]], align 8
-// CHECK1-NEXT:    store i32* [[TMP90]], i32** [[_TMP4]], align 8
-// CHECK1-NEXT:    [[TMP91:%.*]] = load i32*, i32** [[C]], align 8
-// CHECK1-NEXT:    store i32* [[TMP91]], i32** [[_TMP5]], align 8
-// CHECK1-NEXT:    [[TMP92:%.*]] = load %class.anon*, %class.anon** [[L]], align 8
-// CHECK1-NEXT:    store %class.anon* [[TMP92]], %class.anon** [[_TMP6]], align 8
-// CHECK1-NEXT:    [[TMP93:%.*]] = load i32*, i32** [[_TMP4]], align 8
-// CHECK1-NEXT:    [[TMP94:%.*]] = load i32*, i32** [[_TMP5]], align 8
-// CHECK1-NEXT:    [[TMP95:%.*]] = load i32*, i32** [[D]], align 8
-// CHECK1-NEXT:    [[TMP96:%.*]] = load %class.anon*, %class.anon** [[_TMP6]], align 8
-// CHECK1-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP96]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP96]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP99:%.*]] = load i32*, i32** [[TMP98]], align 8
-// CHECK1-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP96]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP96]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP102:%.*]] = load i32*, i32** [[TMP101]], align 8
-// CHECK1-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP96]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP96]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP105:%.*]] = load i32*, i32** [[TMP104]], align 8
-// CHECK1-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP96]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP96]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP108:%.*]] = load i32**, i32*** [[TMP107]], align 8
-// CHECK1-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP96]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP96]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP111:%.*]] = load i32*, i32** [[TMP110]], align 8
-// CHECK1-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i32**
-// CHECK1-NEXT:    store i32* [[ARGC_ADDR]], i32** [[TMP113]], align 8
-// CHECK1-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32**
-// CHECK1-NEXT:    store i32* [[ARGC_ADDR]], i32** [[TMP115]], align 8
-// CHECK1-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP116]], align 8
-// CHECK1-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to i32**
-// CHECK1-NEXT:    store i32* [[TMP93]], i32** [[TMP118]], align 8
-// CHECK1-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to i32**
-// CHECK1-NEXT:    store i32* [[TMP93]], i32** [[TMP120]], align 8
-// CHECK1-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP121]], align 8
-// CHECK1-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP98:%.*]] = load i32*, i32** [[B]], align 8
+// CHECK1-NEXT:    store i32* [[TMP98]], i32** [[_TMP4]], align 8
+// CHECK1-NEXT:    [[TMP99:%.*]] = load i32*, i32** [[C]], align 8
+// CHECK1-NEXT:    store i32* [[TMP99]], i32** [[_TMP5]], align 8
+// CHECK1-NEXT:    [[TMP100:%.*]] = load %class.anon*, %class.anon** [[L]], align 8
+// CHECK1-NEXT:    store %class.anon* [[TMP100]], %class.anon** [[_TMP6]], align 8
+// CHECK1-NEXT:    [[TMP101:%.*]] = load i32*, i32** [[_TMP4]], align 8
+// CHECK1-NEXT:    [[TMP102:%.*]] = load i32*, i32** [[_TMP5]], align 8
+// CHECK1-NEXT:    [[TMP103:%.*]] = load i32*, i32** [[D]], align 8
+// CHECK1-NEXT:    [[TMP104:%.*]] = load %class.anon*, %class.anon** [[_TMP6]], align 8
+// CHECK1-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP104]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP104]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP107:%.*]] = load i32*, i32** [[TMP106]], align 8
+// CHECK1-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP104]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP104]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP110:%.*]] = load i32*, i32** [[TMP109]], align 8
+// CHECK1-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP104]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP104]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP113:%.*]] = load i32*, i32** [[TMP112]], align 8
+// CHECK1-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP104]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP104]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP116:%.*]] = load i32**, i32*** [[TMP115]], align 8
+// CHECK1-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP104]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[CLASS_ANON]], %class.anon* [[TMP104]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP119:%.*]] = load i32*, i32** [[TMP118]], align 8
+// CHECK1-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP121:%.*]] = bitcast i8** [[TMP120]] to i32**
+// CHECK1-NEXT:    store i32* [[ARGC_ADDR]], i32** [[TMP121]], align 8
+// CHECK1-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP123:%.*]] = bitcast i8** [[TMP122]] to i32**
-// CHECK1-NEXT:    store i32* [[TMP94]], i32** [[TMP123]], align 8
-// CHECK1-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP125:%.*]] = bitcast i8** [[TMP124]] to i32**
-// CHECK1-NEXT:    store i32* [[TMP94]], i32** [[TMP125]], align 8
-// CHECK1-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP126]], align 8
-// CHECK1-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i32* [[ARGC_ADDR]], i32** [[TMP123]], align 8
+// CHECK1-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP124]], align 8
+// CHECK1-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i32**
+// CHECK1-NEXT:    store i32* [[TMP101]], i32** [[TMP126]], align 8
+// CHECK1-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
 // CHECK1-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i32**
-// CHECK1-NEXT:    store i32* [[TMP95]], i32** [[TMP128]], align 8
-// CHECK1-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to i32**
-// CHECK1-NEXT:    store i32* [[TMP95]], i32** [[TMP130]], align 8
-// CHECK1-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 3
-// CHECK1-NEXT:    store i8* null, i8** [[TMP131]], align 8
-// CHECK1-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i32* [[TMP101]], i32** [[TMP128]], align 8
+// CHECK1-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP129]], align 8
+// CHECK1-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i32**
+// CHECK1-NEXT:    store i32* [[TMP102]], i32** [[TMP131]], align 8
+// CHECK1-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i32**
-// CHECK1-NEXT:    store i32* [[A]], i32** [[TMP133]], align 8
-// CHECK1-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i32**
-// CHECK1-NEXT:    store i32* [[A]], i32** [[TMP135]], align 8
-// CHECK1-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 4
-// CHECK1-NEXT:    store i8* null, i8** [[TMP136]], align 8
-// CHECK1-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 5
-// CHECK1-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to %class.anon**
-// CHECK1-NEXT:    store %class.anon* [[TMP96]], %class.anon** [[TMP138]], align 8
-// CHECK1-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 5
-// CHECK1-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to %class.anon**
-// CHECK1-NEXT:    store %class.anon* [[TMP96]], %class.anon** [[TMP140]], align 8
-// CHECK1-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 5
-// CHECK1-NEXT:    store i8* null, i8** [[TMP141]], align 8
-// CHECK1-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to i32***
-// CHECK1-NEXT:    store i32** [[TMP97]], i32*** [[TMP143]], align 8
-// CHECK1-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP145:%.*]] = bitcast i8** [[TMP144]] to i32**
-// CHECK1-NEXT:    store i32* [[TMP99]], i32** [[TMP145]], align 8
-// CHECK1-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 6
-// CHECK1-NEXT:    store i8* null, i8** [[TMP146]], align 8
-// CHECK1-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to i32***
-// CHECK1-NEXT:    store i32** [[TMP100]], i32*** [[TMP148]], align 8
-// CHECK1-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP150:%.*]] = bitcast i8** [[TMP149]] to i32**
-// CHECK1-NEXT:    store i32* [[TMP102]], i32** [[TMP150]], align 8
-// CHECK1-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 7
-// CHECK1-NEXT:    store i8* null, i8** [[TMP151]], align 8
-// CHECK1-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to i32***
-// CHECK1-NEXT:    store i32** [[TMP103]], i32*** [[TMP153]], align 8
-// CHECK1-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP155:%.*]] = bitcast i8** [[TMP154]] to i32**
-// CHECK1-NEXT:    store i32* [[TMP105]], i32** [[TMP155]], align 8
-// CHECK1-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 8
-// CHECK1-NEXT:    store i8* null, i8** [[TMP156]], align 8
-// CHECK1-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 9
-// CHECK1-NEXT:    [[TMP158:%.*]] = bitcast i8** [[TMP157]] to i32****
-// CHECK1-NEXT:    store i32*** [[TMP106]], i32**** [[TMP158]], align 8
-// CHECK1-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 9
-// CHECK1-NEXT:    [[TMP160:%.*]] = bitcast i8** [[TMP159]] to i32***
-// CHECK1-NEXT:    store i32** [[TMP108]], i32*** [[TMP160]], align 8
-// CHECK1-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 9
-// CHECK1-NEXT:    store i8* null, i8** [[TMP161]], align 8
-// CHECK1-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 10
-// CHECK1-NEXT:    [[TMP163:%.*]] = bitcast i8** [[TMP162]] to i32***
-// CHECK1-NEXT:    store i32** [[TMP109]], i32*** [[TMP163]], align 8
-// CHECK1-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 10
-// CHECK1-NEXT:    [[TMP165:%.*]] = bitcast i8** [[TMP164]] to i32**
-// CHECK1-NEXT:    store i32* [[TMP111]], i32** [[TMP165]], align 8
-// CHECK1-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 10
-// CHECK1-NEXT:    store i8* null, i8** [[TMP166]], align 8
-// CHECK1-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP169:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l43.region_id, i32 11, i8** [[TMP167]], i8** [[TMP168]], i64* getelementptr inbounds ([11 x i64], [11 x i64]* @.offload_sizes.1, i32 0, i32 0), i64* getelementptr inbounds ([11 x i64], [11 x i64]* @.offload_maptypes.2, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP170:%.*]] = icmp ne i32 [[TMP169]], 0
-// CHECK1-NEXT:    br i1 [[TMP170]], label [[OMP_OFFLOAD_FAILED10:%.*]], label [[OMP_OFFLOAD_CONT11:%.*]]
-// CHECK1:       omp_offload.failed10:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l43(i32* [[ARGC_ADDR]], i32* [[TMP93]], i32* [[TMP94]], i32* [[TMP95]], i32* [[A]], %class.anon* [[TMP96]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT11]]
-// CHECK1:       omp_offload.cont11:
-// CHECK1-NEXT:    [[TMP171:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK1-NEXT:    store i32* [[TMP102]], i32** [[TMP133]], align 8
+// CHECK1-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 2
+// CHECK1-NEXT:    store i8* null, i8** [[TMP134]], align 8
+// CHECK1-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32**
+// CHECK1-NEXT:    store i32* [[TMP103]], i32** [[TMP136]], align 8
+// CHECK1-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i32**
+// CHECK1-NEXT:    store i32* [[TMP103]], i32** [[TMP138]], align 8
+// CHECK1-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 3
+// CHECK1-NEXT:    store i8* null, i8** [[TMP139]], align 8
+// CHECK1-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32**
+// CHECK1-NEXT:    store i32* [[A]], i32** [[TMP141]], align 8
+// CHECK1-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to i32**
+// CHECK1-NEXT:    store i32* [[A]], i32** [[TMP143]], align 8
+// CHECK1-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 4
+// CHECK1-NEXT:    store i8* null, i8** [[TMP144]], align 8
+// CHECK1-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to %class.anon**
+// CHECK1-NEXT:    store %class.anon* [[TMP104]], %class.anon** [[TMP146]], align 8
+// CHECK1-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to %class.anon**
+// CHECK1-NEXT:    store %class.anon* [[TMP104]], %class.anon** [[TMP148]], align 8
+// CHECK1-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 5
+// CHECK1-NEXT:    store i8* null, i8** [[TMP149]], align 8
+// CHECK1-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 6
+// CHECK1-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i32***
+// CHECK1-NEXT:    store i32** [[TMP105]], i32*** [[TMP151]], align 8
+// CHECK1-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 6
+// CHECK1-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to i32**
+// CHECK1-NEXT:    store i32* [[TMP107]], i32** [[TMP153]], align 8
+// CHECK1-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 6
+// CHECK1-NEXT:    store i8* null, i8** [[TMP154]], align 8
+// CHECK1-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP156:%.*]] = bitcast i8** [[TMP155]] to i32***
+// CHECK1-NEXT:    store i32** [[TMP108]], i32*** [[TMP156]], align 8
+// CHECK1-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP158:%.*]] = bitcast i8** [[TMP157]] to i32**
+// CHECK1-NEXT:    store i32* [[TMP110]], i32** [[TMP158]], align 8
+// CHECK1-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 7
+// CHECK1-NEXT:    store i8* null, i8** [[TMP159]], align 8
+// CHECK1-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP161:%.*]] = bitcast i8** [[TMP160]] to i32***
+// CHECK1-NEXT:    store i32** [[TMP111]], i32*** [[TMP161]], align 8
+// CHECK1-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP163:%.*]] = bitcast i8** [[TMP162]] to i32**
+// CHECK1-NEXT:    store i32* [[TMP113]], i32** [[TMP163]], align 8
+// CHECK1-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 8
+// CHECK1-NEXT:    store i8* null, i8** [[TMP164]], align 8
+// CHECK1-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 9
+// CHECK1-NEXT:    [[TMP166:%.*]] = bitcast i8** [[TMP165]] to i32****
+// CHECK1-NEXT:    store i32*** [[TMP114]], i32**** [[TMP166]], align 8
+// CHECK1-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 9
+// CHECK1-NEXT:    [[TMP168:%.*]] = bitcast i8** [[TMP167]] to i32***
+// CHECK1-NEXT:    store i32** [[TMP116]], i32*** [[TMP168]], align 8
+// CHECK1-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 9
+// CHECK1-NEXT:    store i8* null, i8** [[TMP169]], align 8
+// CHECK1-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 10
+// CHECK1-NEXT:    [[TMP171:%.*]] = bitcast i8** [[TMP170]] to i32***
+// CHECK1-NEXT:    store i32** [[TMP117]], i32*** [[TMP171]], align 8
+// CHECK1-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 10
+// CHECK1-NEXT:    [[TMP173:%.*]] = bitcast i8** [[TMP172]] to i32**
+// CHECK1-NEXT:    store i32* [[TMP119]], i32** [[TMP173]], align 8
+// CHECK1-NEXT:    [[TMP174:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 10
+// CHECK1-NEXT:    store i8* null, i8** [[TMP174]], align 8
+// CHECK1-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP176:%.*]] = getelementptr inbounds [11 x i8*], [11 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS10:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP177:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP177]], align 4
+// CHECK1-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 11, i32* [[TMP178]], align 4
+// CHECK1-NEXT:    [[TMP179:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP175]], i8*** [[TMP179]], align 8
+// CHECK1-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP176]], i8*** [[TMP180]], align 8
+// CHECK1-NEXT:    [[TMP181:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([11 x i64], [11 x i64]* @.offload_sizes.1, i32 0, i32 0), i64** [[TMP181]], align 8
+// CHECK1-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([11 x i64], [11 x i64]* @.offload_maptypes.2, i32 0, i32 0), i64** [[TMP182]], align 8
+// CHECK1-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP183]], align 8
+// CHECK1-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP184]], align 8
+// CHECK1-NEXT:    [[TMP185:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l43.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]])
+// CHECK1-NEXT:    [[TMP186:%.*]] = icmp ne i32 [[TMP185]], 0
+// CHECK1-NEXT:    br i1 [[TMP186]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
+// CHECK1:       omp_offload.failed11:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l43(i32* [[ARGC_ADDR]], i32* [[TMP101]], i32* [[TMP102]], i32* [[TMP103]], i32* [[A]], %class.anon* [[TMP104]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
+// CHECK1:       omp_offload.cont12:
+// CHECK1-NEXT:    [[TMP187:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
 // CHECK1-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_ZN1S3fooEv(%struct.S* noundef nonnull align 4 dereferenceable(4) @s)
-// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP171]], [[CALL]]
+// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP187]], [[CALL]]
 // CHECK1-NEXT:    ret i32 [[ADD]]
 //
 //
@@ -540,56 +574,90 @@ int main(int argc, char **argv) {
 // CHECK1-NEXT:    store i8* null, i8** [[TMP19]], align 8
 // CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN1S3fooEv_l27.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.3, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null)
-// CHECK1-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK1-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.3, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN1S3fooEv_l27.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK1-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN1S3fooEv_l27(%struct.S* [[THIS1]], %class.anon.0* [[TMP2]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP24:%.*]] = load %class.anon.0*, %class.anon.0** [[L]], align 8
-// CHECK1-NEXT:    store %class.anon.0* [[TMP24]], %class.anon.0** [[_TMP2]], align 8
-// CHECK1-NEXT:    [[TMP25:%.*]] = load %class.anon.0*, %class.anon.0** [[_TMP2]], align 8
-// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[CLASS_ANON_0]], %class.anon.0* [[TMP25]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[CLASS_ANON_0]], %class.anon.0* [[TMP25]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to %struct.S**
-// CHECK1-NEXT:    store %struct.S* [[THIS1]], %struct.S** [[TMP29]], align 8
-// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to %struct.S**
-// CHECK1-NEXT:    store %struct.S* [[THIS1]], %struct.S** [[TMP31]], align 8
-// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP32]], align 8
-// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP34:%.*]] = bitcast i8** [[TMP33]] to %class.anon.0**
-// CHECK1-NEXT:    store %class.anon.0* [[TMP25]], %class.anon.0** [[TMP34]], align 8
-// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to %class.anon.0**
-// CHECK1-NEXT:    store %class.anon.0* [[TMP25]], %class.anon.0** [[TMP36]], align 8
-// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP37]], align 8
-// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to %struct.S***
-// CHECK1-NEXT:    store %struct.S** [[TMP26]], %struct.S*** [[TMP39]], align 8
-// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to %struct.S***
-// CHECK1-NEXT:    store %struct.S** [[TMP27]], %struct.S*** [[TMP41]], align 8
-// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP42]], align 8
-// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP45:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN1S3fooEv_l29.region_id, i32 3, i8** [[TMP43]], i8** [[TMP44]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.6, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.7, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP46:%.*]] = icmp ne i32 [[TMP45]], 0
-// CHECK1-NEXT:    br i1 [[TMP46]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK1:       omp_offload.failed6:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN1S3fooEv_l29(%struct.S* [[THIS1]], %class.anon.0* [[TMP25]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK1:       omp_offload.cont7:
+// CHECK1-NEXT:    [[TMP32:%.*]] = load %class.anon.0*, %class.anon.0** [[L]], align 8
+// CHECK1-NEXT:    store %class.anon.0* [[TMP32]], %class.anon.0** [[_TMP2]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = load %class.anon.0*, %class.anon.0** [[_TMP2]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[CLASS_ANON_0]], %class.anon.0* [[TMP33]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[CLASS_ANON_0]], %class.anon.0* [[TMP33]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.S**
+// CHECK1-NEXT:    store %struct.S* [[THIS1]], %struct.S** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to %struct.S**
+// CHECK1-NEXT:    store %struct.S* [[THIS1]], %struct.S** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP40]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to %class.anon.0**
+// CHECK1-NEXT:    store %class.anon.0* [[TMP33]], %class.anon.0** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to %class.anon.0**
+// CHECK1-NEXT:    store %class.anon.0* [[TMP33]], %class.anon.0** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to %struct.S***
+// CHECK1-NEXT:    store %struct.S** [[TMP34]], %struct.S*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to %struct.S***
+// CHECK1-NEXT:    store %struct.S** [[TMP35]], %struct.S*** [[TMP49]], align 8
+// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 2
+// CHECK1-NEXT:    store i8* null, i8** [[TMP50]], align 8
+// CHECK1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP53]], align 4
+// CHECK1-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP54]], align 4
+// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP51]], i8*** [[TMP55]], align 8
+// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP52]], i8*** [[TMP56]], align 8
+// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.6, i32 0, i32 0), i64** [[TMP57]], align 8
+// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.7, i32 0, i32 0), i64** [[TMP58]], align 8
+// CHECK1-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP59]], align 8
+// CHECK1-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP60]], align 8
+// CHECK1-NEXT:    [[TMP61:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN1S3fooEv_l29.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK1-NEXT:    [[TMP62:%.*]] = icmp ne i32 [[TMP61]], 0
+// CHECK1-NEXT:    br i1 [[TMP62]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN1S3fooEv_l29(%struct.S* [[THIS1]], %class.anon.0* [[TMP33]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK1:       omp_offload.cont8:
 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP47:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    [[TMP48:%.*]] = load %class.anon.0*, %class.anon.0** [[L]], align 8
-// CHECK1-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z3fooIZN1S3fooEvEUlvE_EiRKT_(%class.anon.0* noundef nonnull align 8 dereferenceable(8) [[TMP48]])
-// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP47]], [[CALL]]
+// CHECK1-NEXT:    [[TMP63:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP64:%.*]] = load %class.anon.0*, %class.anon.0** [[L]], align 8
+// CHECK1-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z3fooIZN1S3fooEvEUlvE_EiRKT_(%class.anon.0* noundef nonnull align 8 dereferenceable(8) [[TMP64]])
+// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP63]], [[CALL]]
 // CHECK1-NEXT:    ret i32 [[ADD]]
 //
 //
@@ -704,9 +772,26 @@ int main(int argc, char **argv) {
 // CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
 // CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooIZN1S3fooEvEUlvE_EiRKT__l18.region_id, i32 2, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.9, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.10, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP16]], align 4
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP17]], align 4
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP14]], i8*** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP15]], i8*** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.9, i32 0, i32 0), i64** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.10, i32 0, i32 0), i64** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP22]], align 8
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooIZN1S3fooEvEUlvE_EiRKT__l18.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP25:%.*]] = icmp ne i32 [[TMP24]], 0
+// CHECK1-NEXT:    br i1 [[TMP25]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooIZN1S3fooEvEUlvE_EiRKT__l18(%class.anon.0* [[TMP1]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/nvptx_lambda_pointer_capturing.cpp b/clang/test/OpenMP/nvptx_lambda_pointer_capturing.cpp
index c02b30d..0d8a614 100644 (file)
--- a/clang/test/OpenMP/nvptx_lambda_pointer_capturing.cpp
+++ b/clang/test/OpenMP/nvptx_lambda_pointer_capturing.cpp
@@ -107,12 +107,10 @@ int main()
 // CHECK: [[CAST_PTR2:%.+]] = bitcast i8** [[LOC_PTR2]] to i32**
 // CHECK: store i32* [[PTR2]], i32** [[CAST_PTR2]]{{.+}}
 
-
-// actual target invocation
-// CHECK: [[BASES_GEP:%.+]] = getelementptr {{.+}} [3 x {{.+}}*], [3 x {{.+}}*]* [[BASE_PTRS]], {{.+}} 0, {{.+}} 0
-// CHECK: [[PTRS_GEP:%.+]] = getelementptr {{.+}} [3 x {{.+}}*], [3 x {{.+}}*]* [[PTRS]], {{.+}} 0, {{.+}} 0
-// CHECK: {{%.+}} = call{{.+}} @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, {{.+}}, {{.+}}, {{.+}}, i8** [[BASES_GEP]], i8** [[PTRS_GEP]], i[[PTRSZ]]* getelementptr inbounds ([3 x i{{.+}}], [3 x i{{.+}}]* [[SIZES]], i{{.+}} 0, i{{.+}} 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* [[TYPES]], i{{.+}} 0, i{{.+}} 0), i8** null, i8** null, {{.+}}, {{.+}})
-
+  // actual target invocation
+  // CHECK: [[BASES_GEP:%.+]] = getelementptr {{.+}} [3 x {{.+}}*], [3 x {{.+}}*]* [[BASE_PTRS]], {{.+}} 0, {{.+}} 0
+  // CHECK: [[PTRS_GEP:%.+]] = getelementptr {{.+}} [3 x {{.+}}*], [3 x {{.+}}*]* [[PTRS]], {{.+}} 0, {{.+}} 0
+  // CHECK: {{%.+}} = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 0, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 
   omp_loop(0,100,body);
   omp_loop_ref(0,100,body);
@@ -150,11 +148,9 @@ int main()
 // CHECK: [[CAST_PTR2:%.+]] = bitcast i8** [[LOC_PTR2]] to i32**
 // CHECK: store i32* [[PTR2]], i32** [[CAST_PTR2]]{{.+}}
 
-
 // actual target invocation
 // CHECK: [[BASES_GEP:%.+]] = getelementptr {{.+}} [5 x {{.+}}*], [5 x {{.+}}*]* [[BASE_PTRS]], {{.+}} 0, {{.+}} 0
 // CHECK: [[PTRS_GEP:%.+]] = getelementptr {{.+}} [5 x {{.+}}*], [5 x {{.+}}*]* [[PTRS]], {{.+}} 0, {{.+}} 0
-// CHECK: {{%.+}} = call{{.+}} @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, {{.+}}, {{.+}}, {{.+}}, i8** [[BASES_GEP]], i8** [[PTRS_GEP]], i[[PTRSZ]]* getelementptr inbounds ([5 x i{{.+}}], [5 x i{{.+}}]* [[SIZES_TEMPLATE]], i{{.+}} 0, i{{.+}} 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* [[TYPES_TEMPLATE]], i{{.+}} 0, i{{.+}} 0), i8** null, i8** null, {{.+}}, {{.+}})
 
 // CHECK: define internal void @{{.+}}omp_loop_ref{{.+}}(
 // CHECK: [[BODY:%body.addr]] = alloca %class.anon*

diff --git a/clang/test/OpenMP/nvptx_target_requires_unified_shared_memory.cpp b/clang/test/OpenMP/nvptx_target_requires_unified_shared_memory.cpp
index bbc4703..a5655b1 100644 (file)
--- a/clang/test/OpenMP/nvptx_target_requires_unified_shared_memory.cpp
+++ b/clang/test/OpenMP/nvptx_target_requires_unified_shared_memory.cpp
@@ -78,7 +78,7 @@ int bar(int n){
 // CHECK-HOST: [[BPTR7:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[OFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK-HOST: [[BPTR8:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[OFFLOAD_PTRS]], i32 0, i32 0
 
-// CHECK-HOST: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.*}}.region_id, i32 2, i8** [[BPTR7]], i8** [[BPTR8]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[OFFLOAD_SIZES]], i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[OFFLOAD_MAPTYPES]], i32 0, i32 0), i8** null, i8** null)
+// CHECK-HOST: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 
 // CHECK-DEVICE: [[VAR_LINK:@.+]] = weak{{.*}} global double* null
 // CHECK-DEVICE: [[VAR_TO:@.+]] = weak{{.*}} global double* null

diff --git a/clang/test/OpenMP/openmp_offload_codegen.cpp b/clang/test/OpenMP/openmp_offload_codegen.cpp
index 918d54a..3409363 100644 (file)
--- a/clang/test/OpenMP/openmp_offload_codegen.cpp
+++ b/clang/test/OpenMP/openmp_offload_codegen.cpp
@@ -35,9 +35,12 @@ void target_maps_parallel_integer(int a){
 // CK1: [[GEPOP:%.+]] = getelementptr inbounds {{.*}}
 // CK1: [[GEPOPBIT:%.+]] = bitcast i8** [[GEPOP]]
 // CK1: store i32* %ParamToKernel, i32** [[GEPOPBIT]]
-// CK1: [[GEPMAPPERARG:%.+]] = getelementptr inbounds {{.*}}
-// CK1: [[GEPOBPARG:%.+]] = getelementptr inbounds {{.*}}
-// CK1: [[GEPOPARG:%.+]] = getelementptr inbounds {{.*}}
-// CK1: call {{.*}}tgt_target_mapper({{.*}}i8** [[GEPOBPARG]], i8** [[GEPOPARG]]{{.*}}, i8** null)
+// CK1: [[GEPOBPARG:%.+]] = getelementptr inbounds {{.*}} %.offload_baseptrs, i32 0, i32 0
+// CK1: [[GEPOPARG:%.+]] = getelementptr inbounds {{.*}} %.offload_ptrs, i32 0, i32 0
+// CK1: [[ARGBP:%.+]] = getelementptr inbounds %struct.__tgt_kernel_arguments, %struct.__tgt_kernel_arguments* %kernel_args, i32 0, i32 2
+// CK1: store i8** [[GEPOBPARG]], i8*** [[ARGBP]], align 8
+// CK1: [[ARGP:%.+]] = getelementptr inbounds %struct.__tgt_kernel_arguments, %struct.__tgt_kernel_arguments* %kernel_args, i32 0, i32 3
+// CK1: store i8** [[GEPOPARG]], i8*** [[ARGP]], align 8
+// CK1: call {{.*}}tgt_target_kernel({{.*}})
 
 #endif

diff --git a/clang/test/OpenMP/reduction_implicit_map.cpp b/clang/test/OpenMP/reduction_implicit_map.cpp
index 8fb6675..06e2495 100644 (file)
--- a/clang/test/OpenMP/reduction_implicit_map.cpp
+++ b/clang/test/OpenMP/reduction_implicit_map.cpp
@@ -335,30 +335,64 @@ int main()
 // CHECK1-NEXT:    store i8* null, i8** [[TMP4]], align 8
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3barv_l50.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3barv_l50.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3barv_l50([5 x %class.S2]* [[O]]) #[[ATTR8:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x [10 x [10 x double]]]**
-// CHECK1-NEXT:    store [10 x [10 x [10 x double]]]* [[B]], [10 x [10 x [10 x double]]]** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x [10 x [10 x double]]]**
-// CHECK1-NEXT:    store [10 x [10 x [10 x double]]]* [[B]], [10 x [10 x [10 x double]]]** [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3barv_l55.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED4:%.*]], label [[OMP_OFFLOAD_CONT5:%.*]]
-// CHECK1:       omp_offload.failed4:
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x [10 x [10 x double]]]**
+// CHECK1-NEXT:    store [10 x [10 x [10 x double]]]* [[B]], [10 x [10 x [10 x double]]]** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x [10 x [10 x double]]]**
+// CHECK1-NEXT:    store [10 x [10 x [10 x double]]]* [[B]], [10 x [10 x [10 x double]]]** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS4:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3barv_l55.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
+// CHECK1:       omp_offload.failed5:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3barv_l55([10 x [10 x [10 x double]]]* [[B]]) #[[ATTR8]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT5]]
-// CHECK1:       omp_offload.cont5:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
+// CHECK1:       omp_offload.cont6:
 // CHECK1-NEXT:    ret i32 0
 //
 //
@@ -827,14 +861,14 @@ int main()
 // CHECK2-NEXT:    [[DOTCAPTURE_EXPR_12:%.*]] = alloca i32, align 4
 // CHECK2-NEXT:    [[DOTCAPTURE_EXPR_13:%.*]] = alloca i32, align 4
 // CHECK2-NEXT:    [[A:%.*]] = alloca [10 x i32], align 4
-// CHECK2-NEXT:    [[SIZE_CASTED20:%.*]] = alloca i32, align 4
-// CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [2 x i8*], align 4
-// CHECK2-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [2 x i8*], align 4
-// CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [2 x i8*], align 4
-// CHECK2-NEXT:    [[SIZE_CASTED27:%.*]] = alloca i32, align 4
-// CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS29:%.*]] = alloca [2 x i8*], align 4
-// CHECK2-NEXT:    [[DOTOFFLOAD_PTRS30:%.*]] = alloca [2 x i8*], align 4
-// CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS31:%.*]] = alloca [2 x i8*], align 4
+// CHECK2-NEXT:    [[SIZE_CASTED21:%.*]] = alloca i32, align 4
+// CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [2 x i8*], align 4
+// CHECK2-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [2 x i8*], align 4
+// CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [2 x i8*], align 4
+// CHECK2-NEXT:    [[SIZE_CASTED29:%.*]] = alloca i32, align 4
+// CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS31:%.*]] = alloca [2 x i8*], align 4
+// CHECK2-NEXT:    [[DOTOFFLOAD_PTRS32:%.*]] = alloca [2 x i8*], align 4
+// CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS33:%.*]] = alloca [2 x i8*], align 4
 // CHECK2-NEXT:    store i32* [[INPUT]], i32** [[INPUT_ADDR]], align 4
 // CHECK2-NEXT:    store i32 [[SIZE]], i32* [[SIZE_ADDR]], align 4
 // CHECK2-NEXT:    store i32* [[OUTPUT]], i32** [[OUTPUT_ADDR]], align 4
@@ -894,134 +928,202 @@ int main()
 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP33]], 1
 // CHECK2-NEXT:    [[TMP34:%.*]] = zext i32 [[ADD]] to i64
 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 [[TMP34]])
-// CHECK2-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l69.region_id, i32 3, i8** [[TMP28]], i8** [[TMP29]], i64* [[TMP30]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK2-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
-// CHECK2-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK2-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK2-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK2-NEXT:    store i32 1, i32* [[TMP35]], align 4
+// CHECK2-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK2-NEXT:    store i32 3, i32* [[TMP36]], align 4
+// CHECK2-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK2-NEXT:    store i8** [[TMP28]], i8*** [[TMP37]], align 4
+// CHECK2-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK2-NEXT:    store i8** [[TMP29]], i8*** [[TMP38]], align 4
+// CHECK2-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK2-NEXT:    store i64* [[TMP30]], i64** [[TMP39]], align 4
+// CHECK2-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK2-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP41]], align 4
+// CHECK2-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK2-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l69.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK2-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
+// CHECK2-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK2:       omp_offload.failed:
 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l69(i32 [[TMP1]], i32* [[TMP2]], i32* [[TMP3]]) #[[ATTR2:[0-9]+]]
 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK2:       omp_offload.cont:
-// CHECK2-NEXT:    [[TMP37:%.*]] = load i32, i32* [[SIZE_ADDR]], align 4
-// CHECK2-NEXT:    store i32 [[TMP37]], i32* [[SIZE_CASTED4]], align 4
-// CHECK2-NEXT:    [[TMP38:%.*]] = load i32, i32* [[SIZE_CASTED4]], align 4
-// CHECK2-NEXT:    [[TMP39:%.*]] = load i32*, i32** [[OUTPUT_ADDR]], align 4
-// CHECK2-NEXT:    [[TMP40:%.*]] = load i32*, i32** [[INPUT_ADDR]], align 4
-// CHECK2-NEXT:    [[TMP41:%.*]] = load i32*, i32** [[OUTPUT_ADDR]], align 4
-// CHECK2-NEXT:    [[TMP42:%.*]] = load i32*, i32** [[OUTPUT_ADDR]], align 4
-// CHECK2-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, i32* [[TMP42]], i32 0
-// CHECK2-NEXT:    [[TMP43:%.*]] = load i32*, i32** [[INPUT_ADDR]], align 4
-// CHECK2-NEXT:    [[TMP44:%.*]] = load i32*, i32** [[INPUT_ADDR]], align 4
-// CHECK2-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i32, i32* [[TMP44]], i32 0
 // CHECK2-NEXT:    [[TMP45:%.*]] = load i32, i32* [[SIZE_ADDR]], align 4
-// CHECK2-NEXT:    [[TMP46:%.*]] = mul nuw i32 [[TMP45]], 4
-// CHECK2-NEXT:    [[TMP47:%.*]] = sext i32 [[TMP46]] to i64
-// CHECK2-NEXT:    [[TMP48:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES10]] to i8*
-// CHECK2-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP48]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.7 to i8*), i32 24, i1 false)
-// CHECK2-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32*
-// CHECK2-NEXT:    store i32 [[TMP38]], i32* [[TMP50]], align 4
-// CHECK2-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i32*
-// CHECK2-NEXT:    store i32 [[TMP38]], i32* [[TMP52]], align 4
-// CHECK2-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK2-NEXT:    store i8* null, i8** [[TMP53]], align 4
-// CHECK2-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK2-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
-// CHECK2-NEXT:    store i32* [[TMP41]], i32** [[TMP55]], align 4
-// CHECK2-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK2-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
-// CHECK2-NEXT:    store i32* [[ARRAYIDX5]], i32** [[TMP57]], align 4
-// CHECK2-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
-// CHECK2-NEXT:    store i8* null, i8** [[TMP58]], align 4
-// CHECK2-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 2
-// CHECK2-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to i32**
-// CHECK2-NEXT:    store i32* [[TMP43]], i32** [[TMP60]], align 4
-// CHECK2-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 2
-// CHECK2-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i32**
-// CHECK2-NEXT:    store i32* [[ARRAYIDX6]], i32** [[TMP62]], align 4
-// CHECK2-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES10]], i32 0, i32 2
-// CHECK2-NEXT:    store i64 [[TMP47]], i64* [[TMP63]], align 4
-// CHECK2-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 2
-// CHECK2-NEXT:    store i8* null, i8** [[TMP64]], align 4
-// CHECK2-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES10]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP68:%.*]] = load i32, i32* [[SIZE_ADDR]], align 4
-// CHECK2-NEXT:    store i32 [[TMP68]], i32* [[DOTCAPTURE_EXPR_12]], align 4
-// CHECK2-NEXT:    [[TMP69:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_12]], align 4
-// CHECK2-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[TMP69]], 0
+// CHECK2-NEXT:    store i32 [[TMP45]], i32* [[SIZE_CASTED4]], align 4
+// CHECK2-NEXT:    [[TMP46:%.*]] = load i32, i32* [[SIZE_CASTED4]], align 4
+// CHECK2-NEXT:    [[TMP47:%.*]] = load i32*, i32** [[OUTPUT_ADDR]], align 4
+// CHECK2-NEXT:    [[TMP48:%.*]] = load i32*, i32** [[INPUT_ADDR]], align 4
+// CHECK2-NEXT:    [[TMP49:%.*]] = load i32*, i32** [[OUTPUT_ADDR]], align 4
+// CHECK2-NEXT:    [[TMP50:%.*]] = load i32*, i32** [[OUTPUT_ADDR]], align 4
+// CHECK2-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, i32* [[TMP50]], i32 0
+// CHECK2-NEXT:    [[TMP51:%.*]] = load i32*, i32** [[INPUT_ADDR]], align 4
+// CHECK2-NEXT:    [[TMP52:%.*]] = load i32*, i32** [[INPUT_ADDR]], align 4
+// CHECK2-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i32, i32* [[TMP52]], i32 0
+// CHECK2-NEXT:    [[TMP53:%.*]] = load i32, i32* [[SIZE_ADDR]], align 4
+// CHECK2-NEXT:    [[TMP54:%.*]] = mul nuw i32 [[TMP53]], 4
+// CHECK2-NEXT:    [[TMP55:%.*]] = sext i32 [[TMP54]] to i64
+// CHECK2-NEXT:    [[TMP56:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES10]] to i8*
+// CHECK2-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP56]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.7 to i8*), i32 24, i1 false)
+// CHECK2-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32*
+// CHECK2-NEXT:    store i32 [[TMP46]], i32* [[TMP58]], align 4
+// CHECK2-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to i32*
+// CHECK2-NEXT:    store i32 [[TMP46]], i32* [[TMP60]], align 4
+// CHECK2-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
+// CHECK2-NEXT:    store i8* null, i8** [[TMP61]], align 4
+// CHECK2-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
+// CHECK2-NEXT:    [[TMP63:%.*]] = bitcast i8** [[TMP62]] to i32**
+// CHECK2-NEXT:    store i32* [[TMP49]], i32** [[TMP63]], align 4
+// CHECK2-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
+// CHECK2-NEXT:    [[TMP65:%.*]] = bitcast i8** [[TMP64]] to i32**
+// CHECK2-NEXT:    store i32* [[ARRAYIDX5]], i32** [[TMP65]], align 4
+// CHECK2-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
+// CHECK2-NEXT:    store i8* null, i8** [[TMP66]], align 4
+// CHECK2-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 2
+// CHECK2-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i32**
+// CHECK2-NEXT:    store i32* [[TMP51]], i32** [[TMP68]], align 4
+// CHECK2-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 2
+// CHECK2-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32**
+// CHECK2-NEXT:    store i32* [[ARRAYIDX6]], i32** [[TMP70]], align 4
+// CHECK2-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES10]], i32 0, i32 2
+// CHECK2-NEXT:    store i64 [[TMP55]], i64* [[TMP71]], align 4
+// CHECK2-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 2
+// CHECK2-NEXT:    store i8* null, i8** [[TMP72]], align 4
+// CHECK2-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES10]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP76:%.*]] = load i32, i32* [[SIZE_ADDR]], align 4
+// CHECK2-NEXT:    store i32 [[TMP76]], i32* [[DOTCAPTURE_EXPR_12]], align 4
+// CHECK2-NEXT:    [[TMP77:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_12]], align 4
+// CHECK2-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[TMP77]], 0
 // CHECK2-NEXT:    [[DIV15:%.*]] = sdiv i32 [[SUB14]], 1
 // CHECK2-NEXT:    [[SUB16:%.*]] = sub nsw i32 [[DIV15]], 1
 // CHECK2-NEXT:    store i32 [[SUB16]], i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK2-NEXT:    [[TMP70:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK2-NEXT:    [[ADD17:%.*]] = add nsw i32 [[TMP70]], 1
-// CHECK2-NEXT:    [[TMP71:%.*]] = zext i32 [[ADD17]] to i64
-// CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 [[TMP71]])
-// CHECK2-NEXT:    [[TMP72:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l73.region_id, i32 3, i8** [[TMP65]], i8** [[TMP66]], i64* [[TMP67]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK2-NEXT:    [[TMP73:%.*]] = icmp ne i32 [[TMP72]], 0
-// CHECK2-NEXT:    br i1 [[TMP73]], label [[OMP_OFFLOAD_FAILED18:%.*]], label [[OMP_OFFLOAD_CONT19:%.*]]
-// CHECK2:       omp_offload.failed18:
-// CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l73(i32 [[TMP38]], i32* [[TMP39]], i32* [[TMP40]]) #[[ATTR2]]
-// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT19]]
-// CHECK2:       omp_offload.cont19:
-// CHECK2-NEXT:    [[TMP74:%.*]] = load i32, i32* [[SIZE_ADDR]], align 4
-// CHECK2-NEXT:    store i32 [[TMP74]], i32* [[SIZE_CASTED20]], align 4
-// CHECK2-NEXT:    [[TMP75:%.*]] = load i32, i32* [[SIZE_CASTED20]], align 4
-// CHECK2-NEXT:    [[ARRAYIDX21:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[A]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to i32*
-// CHECK2-NEXT:    store i32 [[TMP75]], i32* [[TMP77]], align 4
-// CHECK2-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
-// CHECK2-NEXT:    store i32 [[TMP75]], i32* [[TMP79]], align 4
-// CHECK2-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 0
-// CHECK2-NEXT:    store i8* null, i8** [[TMP80]], align 4
-// CHECK2-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
-// CHECK2-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to [10 x i32]**
-// CHECK2-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP82]], align 4
-// CHECK2-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
-// CHECK2-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32**
-// CHECK2-NEXT:    store i32* [[ARRAYIDX21]], i32** [[TMP84]], align 4
-// CHECK2-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 1
-// CHECK2-NEXT:    store i8* null, i8** [[TMP85]], align 4
-// CHECK2-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP88:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l78.region_id, i32 2, i8** [[TMP86]], i8** [[TMP87]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
+// CHECK2-NEXT:    [[TMP78:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
+// CHECK2-NEXT:    [[ADD17:%.*]] = add nsw i32 [[TMP78]], 1
+// CHECK2-NEXT:    [[TMP79:%.*]] = zext i32 [[ADD17]] to i64
+// CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 [[TMP79]])
+// CHECK2-NEXT:    [[KERNEL_ARGS18:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK2-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 0
+// CHECK2-NEXT:    store i32 1, i32* [[TMP80]], align 4
+// CHECK2-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 1
+// CHECK2-NEXT:    store i32 3, i32* [[TMP81]], align 4
+// CHECK2-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 2
+// CHECK2-NEXT:    store i8** [[TMP73]], i8*** [[TMP82]], align 4
+// CHECK2-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 3
+// CHECK2-NEXT:    store i8** [[TMP74]], i8*** [[TMP83]], align 4
+// CHECK2-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 4
+// CHECK2-NEXT:    store i64* [[TMP75]], i64** [[TMP84]], align 4
+// CHECK2-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 5
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP85]], align 4
+// CHECK2-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 6
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP86]], align 4
+// CHECK2-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 7
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP87]], align 4
+// CHECK2-NEXT:    [[TMP88:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l73.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]])
 // CHECK2-NEXT:    [[TMP89:%.*]] = icmp ne i32 [[TMP88]], 0
-// CHECK2-NEXT:    br i1 [[TMP89]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
-// CHECK2:       omp_offload.failed25:
-// CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l78(i32 [[TMP75]], [10 x i32]* [[A]]) #[[ATTR2]]
-// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
-// CHECK2:       omp_offload.cont26:
+// CHECK2-NEXT:    br i1 [[TMP89]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
+// CHECK2:       omp_offload.failed19:
+// CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l73(i32 [[TMP46]], i32* [[TMP47]], i32* [[TMP48]]) #[[ATTR2]]
+// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
+// CHECK2:       omp_offload.cont20:
 // CHECK2-NEXT:    [[TMP90:%.*]] = load i32, i32* [[SIZE_ADDR]], align 4
-// CHECK2-NEXT:    store i32 [[TMP90]], i32* [[SIZE_CASTED27]], align 4
-// CHECK2-NEXT:    [[TMP91:%.*]] = load i32, i32* [[SIZE_CASTED27]], align 4
-// CHECK2-NEXT:    [[ARRAYIDX28:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[A]], i32 0, i32 3
-// CHECK2-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 0
+// CHECK2-NEXT:    store i32 [[TMP90]], i32* [[SIZE_CASTED21]], align 4
+// CHECK2-NEXT:    [[TMP91:%.*]] = load i32, i32* [[SIZE_CASTED21]], align 4
+// CHECK2-NEXT:    [[ARRAYIDX22:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[A]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
 // CHECK2-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to i32*
 // CHECK2-NEXT:    store i32 [[TMP91]], i32* [[TMP93]], align 4
-// CHECK2-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
 // CHECK2-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32*
 // CHECK2-NEXT:    store i32 [[TMP91]], i32* [[TMP95]], align 4
-// CHECK2-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 0
 // CHECK2-NEXT:    store i8* null, i8** [[TMP96]], align 4
-// CHECK2-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 1
+// CHECK2-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
 // CHECK2-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to [10 x i32]**
 // CHECK2-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP98]], align 4
-// CHECK2-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 1
+// CHECK2-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
 // CHECK2-NEXT:    [[TMP100:%.*]] = bitcast i8** [[TMP99]] to i32**
-// CHECK2-NEXT:    store i32* [[ARRAYIDX28]], i32** [[TMP100]], align 4
-// CHECK2-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i32 0, i32 1
+// CHECK2-NEXT:    store i32* [[ARRAYIDX22]], i32** [[TMP100]], align 4
+// CHECK2-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 1
 // CHECK2-NEXT:    store i8* null, i8** [[TMP101]], align 4
-// CHECK2-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP104:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l81.region_id, i32 2, i8** [[TMP102]], i8** [[TMP103]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.15, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.16, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK2-NEXT:    [[TMP105:%.*]] = icmp ne i32 [[TMP104]], 0
-// CHECK2-NEXT:    br i1 [[TMP105]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
-// CHECK2:       omp_offload.failed32:
-// CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l81(i32 [[TMP91]], [10 x i32]* [[A]]) #[[ATTR2]]
-// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
-// CHECK2:       omp_offload.cont33:
+// CHECK2-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
+// CHECK2-NEXT:    [[KERNEL_ARGS26:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK2-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 0
+// CHECK2-NEXT:    store i32 1, i32* [[TMP104]], align 4
+// CHECK2-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 1
+// CHECK2-NEXT:    store i32 2, i32* [[TMP105]], align 4
+// CHECK2-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 2
+// CHECK2-NEXT:    store i8** [[TMP102]], i8*** [[TMP106]], align 4
+// CHECK2-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 3
+// CHECK2-NEXT:    store i8** [[TMP103]], i8*** [[TMP107]], align 4
+// CHECK2-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 4
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP108]], align 4
+// CHECK2-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 5
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP109]], align 4
+// CHECK2-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 6
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP110]], align 4
+// CHECK2-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 7
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP111]], align 4
+// CHECK2-NEXT:    [[TMP112:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l78.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]])
+// CHECK2-NEXT:    [[TMP113:%.*]] = icmp ne i32 [[TMP112]], 0
+// CHECK2-NEXT:    br i1 [[TMP113]], label [[OMP_OFFLOAD_FAILED27:%.*]], label [[OMP_OFFLOAD_CONT28:%.*]]
+// CHECK2:       omp_offload.failed27:
+// CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l78(i32 [[TMP91]], [10 x i32]* [[A]]) #[[ATTR2]]
+// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT28]]
+// CHECK2:       omp_offload.cont28:
+// CHECK2-NEXT:    [[TMP114:%.*]] = load i32, i32* [[SIZE_ADDR]], align 4
+// CHECK2-NEXT:    store i32 [[TMP114]], i32* [[SIZE_CASTED29]], align 4
+// CHECK2-NEXT:    [[TMP115:%.*]] = load i32, i32* [[SIZE_CASTED29]], align 4
+// CHECK2-NEXT:    [[ARRAYIDX30:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[A]], i32 0, i32 3
+// CHECK2-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS31]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i32*
+// CHECK2-NEXT:    store i32 [[TMP115]], i32* [[TMP117]], align 4
+// CHECK2-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS32]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP119:%.*]] = bitcast i8** [[TMP118]] to i32*
+// CHECK2-NEXT:    store i32 [[TMP115]], i32* [[TMP119]], align 4
+// CHECK2-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS33]], i32 0, i32 0
+// CHECK2-NEXT:    store i8* null, i8** [[TMP120]], align 4
+// CHECK2-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS31]], i32 0, i32 1
+// CHECK2-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to [10 x i32]**
+// CHECK2-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP122]], align 4
+// CHECK2-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS32]], i32 0, i32 1
+// CHECK2-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to i32**
+// CHECK2-NEXT:    store i32* [[ARRAYIDX30]], i32** [[TMP124]], align 4
+// CHECK2-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS33]], i32 0, i32 1
+// CHECK2-NEXT:    store i8* null, i8** [[TMP125]], align 4
+// CHECK2-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS31]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS32]], i32 0, i32 0
+// CHECK2-NEXT:    [[KERNEL_ARGS34:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK2-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 0
+// CHECK2-NEXT:    store i32 1, i32* [[TMP128]], align 4
+// CHECK2-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 1
+// CHECK2-NEXT:    store i32 2, i32* [[TMP129]], align 4
+// CHECK2-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 2
+// CHECK2-NEXT:    store i8** [[TMP126]], i8*** [[TMP130]], align 4
+// CHECK2-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 3
+// CHECK2-NEXT:    store i8** [[TMP127]], i8*** [[TMP131]], align 4
+// CHECK2-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 4
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.15, i32 0, i32 0), i64** [[TMP132]], align 4
+// CHECK2-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 5
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.16, i32 0, i32 0), i64** [[TMP133]], align 4
+// CHECK2-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 6
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP134]], align 4
+// CHECK2-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 7
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP135]], align 4
+// CHECK2-NEXT:    [[TMP136:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]])
+// CHECK2-NEXT:    [[TMP137:%.*]] = icmp ne i32 [[TMP136]], 0
+// CHECK2-NEXT:    br i1 [[TMP137]], label [[OMP_OFFLOAD_FAILED35:%.*]], label [[OMP_OFFLOAD_CONT36:%.*]]
+// CHECK2:       omp_offload.failed35:
+// CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3sumPiiS__l81(i32 [[TMP115]], [10 x i32]* [[A]]) #[[ATTR2]]
+// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT36]]
+// CHECK2:       omp_offload.cont36:
 // CHECK2-NEXT:    ret void
 //
 //

diff --git a/clang/test/OpenMP/target_codegen.cpp b/clang/test/OpenMP/target_codegen.cpp
index f73856f..a663175 100644 (file)
--- a/clang/test/OpenMP/target_codegen.cpp
+++ b/clang/test/OpenMP/target_codegen.cpp
@@ -134,18 +134,18 @@ int foo(int n) {
   TT<long long, char> d;
   static long *plocal;
 
-  // CHECK:       [[ADD:%.+]] = add nsw i32
-  // CHECK:       store i32 [[ADD]], i32* [[DEVICE_CAP:%.+]],
-  // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
-  // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-  // CHECK:       [[RET:%.+]] = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null)
-  // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
-  // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
-  // CHECK:       [[FAIL]]
-  // CHECK:       call void [[HVT0:@.+]]()
-  // CHECK-NEXT:  br label %[[END]]
-  // CHECK:       [[END]]
-  #pragma omp target device(global + a)
+// CHECK:       [[ADD:%.+]] = add nsw i32
+// CHECK:       store i32 [[ADD]], i32* [[DEVICE_CAP:%.+]],
+// CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
+// CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 {{.+}}, i32 {{.+}}, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
+// CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
+// CHECK:       [[FAIL]]
+// CHECK:       call void [[HVT0:@.+]]()
+// CHECK-NEXT:  br label %[[END]]
+// CHECK:       [[END]]
+#pragma omp target device(global + a)
   {
   }
 
@@ -194,60 +194,68 @@ int foo(int n) {
     global += 1;
   }
 
-  // CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 1, i8** [[BP:%[^,]+]], i8** [[P:%[^,]+]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZET2]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPT2]], i32 0, i32 0), i8** null, i8** null)
-  // CHECK-DAG:   [[BP]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[BPR:%[^,]+]], i32 0, i32 0
-  // CHECK-DAG:   [[P]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[PR:%[^,]+]], i32 0, i32 0
-  // CHECK-DAG:   [[BPADDR0:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[BPR]], i32 0, i32 [[IDX0:[0-9]+]]
-  // CHECK-DAG:   [[PADDR0:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[PR]], i32 0, i32 [[IDX0]]
-  // CHECK-DAG:   [[CBPADDR0:%.+]] = bitcast i8** [[BPADDR0]] to i[[SZ]]*
-  // CHECK-DAG:   [[CPADDR0:%.+]] = bitcast i8** [[PADDR0]] to i[[SZ]]*
-  // CHECK-DAG:   store i[[SZ]] [[BP0:%[^,]+]], i[[SZ]]* [[CBPADDR0]]
-  // CHECK-DAG:   store i[[SZ]] [[P0:%[^,]+]], i[[SZ]]* [[CPADDR0]]
-
-  // CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
-  // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
-  // CHECK:       [[FAIL]]
-  // CHECK:       call void [[HVT2:@.+]](i[[SZ]] {{[^,]+}})
-  // CHECK-NEXT:  br label %[[END]]
-  // CHECK:       [[END]]
-  #pragma omp target if(1)
+// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 {{.+}}, i32 {{.+}}, i32 {{.+}}, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CHECK-DAG:   [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CHECK-DAG:   store i8** [[BP:%.+]], i8*** [[BPARG]]
+// CHECK-DAG:   [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CHECK-DAG:   store i8** [[P:%.+]], i8*** [[PARG]]
+// CHECK-DAG:   [[BP]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[BPR:%[^,]+]], i32 0, i32 0
+// CHECK-DAG:   [[P]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[PR:%[^,]+]], i32 0, i32 0
+// CHECK-DAG:   [[BPADDR0:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[BPR]], i32 0, i32 [[IDX0:[0-9]+]]
+// CHECK-DAG:   [[PADDR0:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[PR]], i32 0, i32 [[IDX0]]
+// CHECK-DAG:   [[CBPADDR0:%.+]] = bitcast i8** [[BPADDR0]] to i[[SZ]]*
+// CHECK-DAG:   [[CPADDR0:%.+]] = bitcast i8** [[PADDR0]] to i[[SZ]]*
+// CHECK-DAG:   store i[[SZ]] [[BP0:%[^,]+]], i[[SZ]]* [[CBPADDR0]]
+// CHECK-DAG:   store i[[SZ]] [[P0:%[^,]+]], i[[SZ]]* [[CPADDR0]]
+
+// CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
+// CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
+// CHECK:       [[FAIL]]
+// CHECK:       call void [[HVT2:@.+]](i[[SZ]] {{[^,]+}})
+// CHECK-NEXT:  br label %[[END]]
+// CHECK:       [[END]]
+#pragma omp target if (1)
   {
     aa += 1;
   }
 
-  // CHECK:       [[IF:%.+]] = icmp sgt i32 {{[^,]+}}, 10
-  // CHECK:       br i1 [[IF]], label %[[IFTHEN:[^,]+]], label %[[IFELSE:[^,]+]]
-  // CHECK:       [[IFTHEN]]
-  // CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 2, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[SIZET3]], i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT3]], i32 0, i32 0), i8** null, i8** null)
-  // CHECK-DAG:   [[BPR]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[BP:%[^,]+]], i32 0, i32 0
-  // CHECK-DAG:   [[PR]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[P:%[^,]+]], i32 0, i32 0
-
-  // CHECK-DAG:   [[BPADDR0:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[BP]], i32 0, i32 0
-  // CHECK-DAG:   [[PADDR0:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[P]], i32 0, i32 0
-  // CHECK-DAG:   [[CBPADDR0:%.+]] = bitcast i8** [[BPADDR0]] to i[[SZ]]*
-  // CHECK-DAG:   [[CPADDR0:%.+]] = bitcast i8** [[PADDR0]] to i[[SZ]]*
-  // CHECK-DAG:   store i[[SZ]] [[BP0:%[^,]+]], i[[SZ]]* [[CBPADDR0]]
-  // CHECK-DAG:   store i[[SZ]] [[P0:%[^,]+]], i[[SZ]]* [[CPADDR0]]
-
-  // CHECK-DAG:   [[BPADDR1:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[BP]], i32 0, i32 1
-  // CHECK-DAG:   [[PADDR1:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[P]], i32 0, i32 1
-  // CHECK-DAG:   [[CBPADDR1:%.+]] = bitcast i8** [[BPADDR1]] to i[[SZ]]*
-  // CHECK-DAG:   [[CPADDR1:%.+]] = bitcast i8** [[PADDR1]] to i[[SZ]]*
-  // CHECK-DAG:   store i[[SZ]] [[BP1:%[^,]+]], i[[SZ]]* [[CBPADDR1]]
-  // CHECK-DAG:   store i[[SZ]] [[P1:%[^,]+]], i[[SZ]]* [[CPADDR1]]
-  // CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
-  // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:.+]], label %[[END:[^,]+]]
-  // CHECK:       [[FAIL]]
-  // CHECK:       call void [[HVT3:@.+]]({{[^,]+}}, {{[^,]+}})
-  // CHECK-NEXT:  br label %[[END]]
-  // CHECK:       [[END]]
-  // CHECK-NEXT:  br label %[[IFEND:.+]]
-  // CHECK:       [[IFELSE]]
-  // CHECK:       call void [[HVT3]]({{[^,]+}}, {{[^,]+}})
-  // CHECK-NEXT:  br label %[[IFEND]]
-
-  // CHECK:       [[IFEND]]
-  #pragma omp target if(n>10)
+// CHECK:       [[IF:%.+]] = icmp sgt i32 {{[^,]+}}, 10
+// CHECK:       br i1 [[IF]], label %[[IFTHEN:[^,]+]], label %[[IFELSE:[^,]+]]
+// CHECK:       [[IFTHEN]]
+// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 {{.+}}, i32 {{.+}}, i32 {{.+}}, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CHECK-DAG:   [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CHECK-DAG:   store i8** [[BPR:%.+]], i8*** [[BPARG]]
+// CHECK-DAG:   [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CHECK-DAG:   store i8** [[PR:%.+]], i8*** [[PARG]]
+// CHECK-DAG:   [[BPR]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[BP:%[^,]+]], i32 0, i32 0
+// CHECK-DAG:   [[PR]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[P:%[^,]+]], i32 0, i32 0
+
+// CHECK-DAG:   [[BPADDR0:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[BP]], i32 0, i32 0
+// CHECK-DAG:   [[PADDR0:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[P]], i32 0, i32 0
+// CHECK-DAG:   [[CBPADDR0:%.+]] = bitcast i8** [[BPADDR0]] to i[[SZ]]*
+// CHECK-DAG:   [[CPADDR0:%.+]] = bitcast i8** [[PADDR0]] to i[[SZ]]*
+// CHECK-DAG:   store i[[SZ]] [[BP0:%[^,]+]], i[[SZ]]* [[CBPADDR0]]
+// CHECK-DAG:   store i[[SZ]] [[P0:%[^,]+]], i[[SZ]]* [[CPADDR0]]
+
+// CHECK-DAG:   [[BPADDR1:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[BP]], i32 0, i32 1
+// CHECK-DAG:   [[PADDR1:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[P]], i32 0, i32 1
+// CHECK-DAG:   [[CBPADDR1:%.+]] = bitcast i8** [[BPADDR1]] to i[[SZ]]*
+// CHECK-DAG:   [[CPADDR1:%.+]] = bitcast i8** [[PADDR1]] to i[[SZ]]*
+// CHECK-DAG:   store i[[SZ]] [[BP1:%[^,]+]], i[[SZ]]* [[CBPADDR1]]
+// CHECK-DAG:   store i[[SZ]] [[P1:%[^,]+]], i[[SZ]]* [[CPADDR1]]
+// CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
+// CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:.+]], label %[[END:[^,]+]]
+// CHECK:       [[FAIL]]
+// CHECK:       call void [[HVT3:@.+]]({{[^,]+}}, {{[^,]+}})
+// CHECK-NEXT:  br label %[[END]]
+// CHECK:       [[END]]
+// CHECK-NEXT:  br label %[[IFEND:.+]]
+// CHECK:       [[IFELSE]]
+// CHECK:       call void [[HVT3]]({{[^,]+}}, {{[^,]+}})
+// CHECK-NEXT:  br label %[[IFEND]]
+
+// CHECK:       [[IFEND]]
+#pragma omp target if (n > 10)
   {
     a += 1;
     aa += 1;
@@ -274,94 +282,100 @@ int foo(int n) {
   // CHECK-32:    [[CNSZSIZE:%.+]] = mul nuw i32 [[CNELEMSIZE2]], 8
   // CHECK-32:    [[CNSIZE:%.+]] = sext i32 [[CNSZSIZE]] to i64
 
-  // CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 9, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SZ4:%[^,]+]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* [[MAPT4]], i32 0, i32 0), i8** null, i8** null)
-  // CHECK-DAG:   [[BPR]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP:%[^,]+]], i32 0, i32 0
-  // CHECK-DAG:   [[PR]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P:%[^,]+]], i32 0, i32 0
-  // CHECK-DAG:   [[SZ4]] = getelementptr inbounds [9 x i64], [9 x i64]* [[PSZ:%[^,]+]], i32 0, i32 0
-
-  // CHECK-DAG:   [[BPADDR0:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX0:0]]
-  // CHECK-DAG:   [[PADDR0:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX0]]
-  // CHECK-DAG:   [[BPADDR1:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX1:1]]
-  // CHECK-DAG:   [[PADDR1:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX1]]
-  // CHECK-DAG:   [[BPADDR2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX2:2]]
-  // CHECK-DAG:   [[PADDR2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX2]]
-  // CHECK-DAG:   [[BPADDR3:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX3:3]]
-  // CHECK-DAG:   [[PADDR3:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX3]]
-  // CHECK-DAG:   [[PSZ3:%.+]] = getelementptr inbounds [9 x i64], [9 x i64]* [[PSZ]], i32 0, i32 [[IDX3]]
-  // CHECK-DAG:   [[BPADDR4:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX4:4]]
-  // CHECK-DAG:   [[PADDR4:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX4]]
-  // CHECK-DAG:   [[BPADDR5:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX5:5]]
-  // CHECK-DAG:   [[PADDR5:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX5]]
-  // CHECK-DAG:   [[BPADDR6:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX6:6]]
-  // CHECK-DAG:   [[PADDR6:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX6]]
-  // CHECK-DAG:   [[BPADDR7:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX7:7]]
-  // CHECK-DAG:   [[PADDR7:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX7]]
-  // CHECK-DAG:   [[PSZ7:%.+]] = getelementptr inbounds [9 x i64], [9 x i64]* [[PSZ]], i32 0, i32 [[IDX7]]
-  // CHECK-DAG:   [[BPADDR8:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX8:8]]
-  // CHECK-DAG:   [[PADDR8:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX8]]
-
-  // The names below are not necessarily consistent with the names used for the
-  // addresses above as some are repeated.
-  // CHECK-DAG:   [[CBPADDR2:%.+]] = bitcast i8** [[BPADDR2]] to i[[SZ]]*
-  // CHECK-DAG:   [[CPADDR2:%.+]] = bitcast i8** [[PADDR2]] to i[[SZ]]*
-  // CHECK-DAG:   store i[[SZ]] [[VLA0]], i[[SZ]]* [[CBPADDR2]]
-  // CHECK-DAG:   store i[[SZ]] [[VLA0]], i[[SZ]]* [[CPADDR2]]
-
-  // CHECK-DAG:   [[CBPADDR6:%.+]] = bitcast i8** [[BPADDR6]] to i[[SZ]]*
-  // CHECK-DAG:   [[CPADDR6:%.+]] = bitcast i8** [[PADDR6]] to i[[SZ]]*
-  // CHECK-DAG:   store i[[SZ]] [[VLA1]], i[[SZ]]* [[CBPADDR6]]
-  // CHECK-DAG:   store i[[SZ]] [[VLA1]], i[[SZ]]* [[CPADDR6]]
-
-  // CHECK-DAG:   [[CBPADDR5:%.+]] = bitcast i8** [[BPADDR5]] to i[[SZ]]*
-  // CHECK-DAG:   [[CPADDR5:%.+]] = bitcast i8** [[PADDR5]] to i[[SZ]]*
-  // CHECK-DAG:   store i[[SZ]] 5, i[[SZ]]* [[CBPADDR5]]
-  // CHECK-DAG:   store i[[SZ]] 5, i[[SZ]]* [[CPADDR5]]
-
-  // CHECK-DAG:   [[CBPADDR0:%.+]] = bitcast i8** [[BPADDR0]] to i[[SZ]]*
-  // CHECK-DAG:   [[CPADDR0:%.+]] = bitcast i8** [[PADDR0]] to i[[SZ]]*
-  // CHECK-DAG:   store i[[SZ]] [[A_CVAL]], i[[SZ]]* [[CBPADDR0]]
-  // CHECK-DAG:   store i[[SZ]] [[A_CVAL]], i[[SZ]]* [[CPADDR0]]
-
-  // CHECK-DAG:   [[CBPADDR1:%.+]] = bitcast i8** [[BPADDR1]] to [10 x float]**
-  // CHECK-DAG:   [[CPADDR1:%.+]] = bitcast i8** [[PADDR1]] to [10 x float]**
-  // CHECK-DAG:   store [10 x float]* %{{.+}}, [10 x float]** [[CBPADDR1]]
-  // CHECK-DAG:   store [10 x float]* %{{.+}}, [10 x float]** [[CPADDR1]]
-
-  // CHECK-DAG:   [[CBPADDR3:%.+]] = bitcast i8** [[BPADDR3]] to float**
-  // CHECK-DAG:   [[CPADDR3:%.+]] = bitcast i8** [[PADDR3]] to float**
-  // CHECK-DAG:   store float* %{{.+}}, float** [[CBPADDR3]]
-  // CHECK-DAG:   store float* %{{.+}}, float** [[CPADDR3]]
-  // CHECK-DAG:   store i64 [[BNSIZE]], i64* [[PSZ3]]
-
-  // CHECK-DAG:   [[CBPADDR4:%.+]] = bitcast i8** [[BPADDR4]] to [5 x [10 x double]]**
-  // CHECK-DAG:   [[CPADDR4:%.+]] = bitcast i8** [[PADDR4]] to [5 x [10 x double]]**
-  // CHECK-DAG:   store [5 x [10 x double]]* %{{.+}}, [5 x [10 x double]]** [[CBPADDR4]]
-  // CHECK-DAG:   store [5 x [10 x double]]* %{{.+}}, [5 x [10 x double]]** [[CPADDR4]]
-
-  // CHECK-DAG:   [[CBPADDR7:%.+]] = bitcast i8** [[BPADDR7]] to double**
-  // CHECK-DAG:   [[CPADDR7:%.+]] = bitcast i8** [[PADDR7]] to double**
-  // CHECK-DAG:   store double* %{{.+}}, double** [[CBPADDR7]]
-  // CHECK-DAG:   store double* %{{.+}}, double** [[CPADDR7]]
-  // CHECK-DAG:   store i64 [[CNSIZE]], i64* [[PSZ7]]
-
-  // CHECK-DAG:   [[CBPADDR8:%.+]] = bitcast i8** [[BPADDR8]] to [[TT]]**
-  // CHECK-DAG:   [[CPADDR8:%.+]] = bitcast i8** [[PADDR8]] to [[TT]]**
-  // CHECK-DAG:   store [[TT]]* %{{.+}}, [[TT]]** [[CBPADDR8]]
-  // CHECK-DAG:   store [[TT]]* %{{.+}}, [[TT]]** [[CPADDR8]]
-
-  // CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
-  // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:.+]], label %[[END:[^,]+]]
-  // CHECK:       [[FAIL]]
-  // CHECK:       call void [[HVT4:@.+]]({{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}})
-  // CHECK-NEXT:  br label %[[END]]
-  // CHECK:       [[END]]
-  // CHECK-NEXT:  br label %[[IFEND:.+]]
-  // CHECK:       [[IFELSE]]
-  // CHECK:       call void [[HVT4]]({{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}})
-  // CHECK-NEXT:  br label %[[IFEND]]
-
-  // CHECK:       [[IFEND]]
-  #pragma omp target if(n>20)
+// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 {{.+}}, i32 {{.+}}, i32 {{.+}}, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CHECK-DAG:   [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CHECK-DAG:   store i8** [[BPR:%.+]], i8*** [[BPARG]]
+// CHECK-DAG:   [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CHECK-DAG:   store i8** [[PR:%.+]], i8*** [[PARG]]
+// CHECK-DAG:   [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CHECK-DAG:   store i64* [[SZ4:%.+]], i64** [[SARG]]
+// CHECK-DAG:   [[BPR]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP:%[^,]+]], i32 0, i32 0
+// CHECK-DAG:   [[PR]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P:%[^,]+]], i32 0, i32 0
+// CHECK-DAG:   [[SZ4]] = getelementptr inbounds [9 x i64], [9 x i64]* [[PSZ:%[^,]+]], i32 0, i32 0
+
+// CHECK-DAG:   [[BPADDR0:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX0:0]]
+// CHECK-DAG:   [[PADDR0:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX0]]
+// CHECK-DAG:   [[BPADDR1:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX1:1]]
+// CHECK-DAG:   [[PADDR1:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX1]]
+// CHECK-DAG:   [[BPADDR2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX2:2]]
+// CHECK-DAG:   [[PADDR2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX2]]
+// CHECK-DAG:   [[BPADDR3:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX3:3]]
+// CHECK-DAG:   [[PADDR3:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX3]]
+// CHECK-DAG:   [[PSZ3:%.+]] = getelementptr inbounds [9 x i64], [9 x i64]* [[PSZ]], i32 0, i32 [[IDX3]]
+// CHECK-DAG:   [[BPADDR4:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX4:4]]
+// CHECK-DAG:   [[PADDR4:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX4]]
+// CHECK-DAG:   [[BPADDR5:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX5:5]]
+// CHECK-DAG:   [[PADDR5:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX5]]
+// CHECK-DAG:   [[BPADDR6:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX6:6]]
+// CHECK-DAG:   [[PADDR6:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX6]]
+// CHECK-DAG:   [[BPADDR7:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX7:7]]
+// CHECK-DAG:   [[PADDR7:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX7]]
+// CHECK-DAG:   [[PSZ7:%.+]] = getelementptr inbounds [9 x i64], [9 x i64]* [[PSZ]], i32 0, i32 [[IDX7]]
+// CHECK-DAG:   [[BPADDR8:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX8:8]]
+// CHECK-DAG:   [[PADDR8:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX8]]
+
+// The names below are not necessarily consistent with the names used for the
+// addresses above as some are repeated.
+// CHECK-DAG:   [[CBPADDR2:%.+]] = bitcast i8** [[BPADDR2]] to i[[SZ]]*
+// CHECK-DAG:   [[CPADDR2:%.+]] = bitcast i8** [[PADDR2]] to i[[SZ]]*
+// CHECK-DAG:   store i[[SZ]] [[VLA0]], i[[SZ]]* [[CBPADDR2]]
+// CHECK-DAG:   store i[[SZ]] [[VLA0]], i[[SZ]]* [[CPADDR2]]
+
+// CHECK-DAG:   [[CBPADDR6:%.+]] = bitcast i8** [[BPADDR6]] to i[[SZ]]*
+// CHECK-DAG:   [[CPADDR6:%.+]] = bitcast i8** [[PADDR6]] to i[[SZ]]*
+// CHECK-DAG:   store i[[SZ]] [[VLA1]], i[[SZ]]* [[CBPADDR6]]
+// CHECK-DAG:   store i[[SZ]] [[VLA1]], i[[SZ]]* [[CPADDR6]]
+
+// CHECK-DAG:   [[CBPADDR5:%.+]] = bitcast i8** [[BPADDR5]] to i[[SZ]]*
+// CHECK-DAG:   [[CPADDR5:%.+]] = bitcast i8** [[PADDR5]] to i[[SZ]]*
+// CHECK-DAG:   store i[[SZ]] 5, i[[SZ]]* [[CBPADDR5]]
+// CHECK-DAG:   store i[[SZ]] 5, i[[SZ]]* [[CPADDR5]]
+
+// CHECK-DAG:   [[CBPADDR0:%.+]] = bitcast i8** [[BPADDR0]] to i[[SZ]]*
+// CHECK-DAG:   [[CPADDR0:%.+]] = bitcast i8** [[PADDR0]] to i[[SZ]]*
+// CHECK-DAG:   store i[[SZ]] [[A_CVAL]], i[[SZ]]* [[CBPADDR0]]
+// CHECK-DAG:   store i[[SZ]] [[A_CVAL]], i[[SZ]]* [[CPADDR0]]
+
+// CHECK-DAG:   [[CBPADDR1:%.+]] = bitcast i8** [[BPADDR1]] to [10 x float]**
+// CHECK-DAG:   [[CPADDR1:%.+]] = bitcast i8** [[PADDR1]] to [10 x float]**
+// CHECK-DAG:   store [10 x float]* %{{.+}}, [10 x float]** [[CBPADDR1]]
+// CHECK-DAG:   store [10 x float]* %{{.+}}, [10 x float]** [[CPADDR1]]
+
+// CHECK-DAG:   [[CBPADDR3:%.+]] = bitcast i8** [[BPADDR3]] to float**
+// CHECK-DAG:   [[CPADDR3:%.+]] = bitcast i8** [[PADDR3]] to float**
+// CHECK-DAG:   store float* %{{.+}}, float** [[CBPADDR3]]
+// CHECK-DAG:   store float* %{{.+}}, float** [[CPADDR3]]
+// CHECK-DAG:   store i64 [[BNSIZE]], i64* [[PSZ3]]
+
+// CHECK-DAG:   [[CBPADDR4:%.+]] = bitcast i8** [[BPADDR4]] to [5 x [10 x double]]**
+// CHECK-DAG:   [[CPADDR4:%.+]] = bitcast i8** [[PADDR4]] to [5 x [10 x double]]**
+// CHECK-DAG:   store [5 x [10 x double]]* %{{.+}}, [5 x [10 x double]]** [[CBPADDR4]]
+// CHECK-DAG:   store [5 x [10 x double]]* %{{.+}}, [5 x [10 x double]]** [[CPADDR4]]
+
+// CHECK-DAG:   [[CBPADDR7:%.+]] = bitcast i8** [[BPADDR7]] to double**
+// CHECK-DAG:   [[CPADDR7:%.+]] = bitcast i8** [[PADDR7]] to double**
+// CHECK-DAG:   store double* %{{.+}}, double** [[CBPADDR7]]
+// CHECK-DAG:   store double* %{{.+}}, double** [[CPADDR7]]
+// CHECK-DAG:   store i64 [[CNSIZE]], i64* [[PSZ7]]
+
+// CHECK-DAG:   [[CBPADDR8:%.+]] = bitcast i8** [[BPADDR8]] to [[TT]]**
+// CHECK-DAG:   [[CPADDR8:%.+]] = bitcast i8** [[PADDR8]] to [[TT]]**
+// CHECK-DAG:   store [[TT]]* %{{.+}}, [[TT]]** [[CBPADDR8]]
+// CHECK-DAG:   store [[TT]]* %{{.+}}, [[TT]]** [[CPADDR8]]
+
+// CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
+// CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:.+]], label %[[END:[^,]+]]
+// CHECK:       [[FAIL]]
+// CHECK:       call void [[HVT4:@.+]]({{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}})
+// CHECK-NEXT:  br label %[[END]]
+// CHECK:       [[END]]
+// CHECK-NEXT:  br label %[[IFEND:.+]]
+// CHECK:       [[IFELSE]]
+// CHECK:       call void [[HVT4]]({{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}})
+// CHECK-NEXT:  br label %[[IFEND]]
+
+// CHECK:       [[IFEND]]
+#pragma omp target if (n > 20)
   {
     a += 1;
     b[2] += 1.0;
@@ -382,7 +396,13 @@ int foo(int n) {
 
 // CHECK: define internal void [[HVT0_:@.+]](i[[SZ]]* noundef {{%[^,]+}}, i[[SZ]] noundef {{%[^,]+}})
 // CHECK: define internal {{.*}}i32 [[OMP_TASK_ENTRY]](i32 {{.*}}%0, [[KMP_TASK_T_WITH_PRIVATES]]* noalias noundef %1)
-// CHECK-DAG: [[RET:%.+]] = call i32 @__tgt_target_nowait_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE:%.+]], i8* @{{[^,]+}}, i32 2, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SIZE:%.+]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT]], i32 0, i32 0), i8** null, i8** null, i32 0, i8* null, i32 0, i8* null)
+// CHECK-DAG: [[RET:%.+]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 [[DEVICE:%.+]], i32 {{.+}}, i32 {{.+}}, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]], i32 0, i8* null, i32 0, i8* null)
+// CHECK-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CHECK-DAG: store i8** [[BPR:%.+]], i8*** [[BPARG]]
+// CHECK-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CHECK-DAG: store i8** [[PR:%.+]], i8*** [[PARG]]
+// CHECK-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CHECK-DAG: store i64* [[SIZE:%.+]], i64** [[SARG]]
 // CHECK-DAG: [[DEVICE]] = sext i32 [[DEV:%.+]] to i64
 // CHECK-DAG: [[DEV]] = load i32, i32* [[DEVADDR:%.+]], align
 // CHECK-DAG: [[DEVADDR]] = getelementptr inbounds [[ANON_T]], [[ANON_T]]* {{%.+}}, i32 0, i32 2
@@ -568,7 +588,13 @@ int bar(int n){
 // CHECK-32:    [[CSZSIZE:%.+]] = mul nuw i32 [[CELEMSIZE2]], 2
 // CHECK-32:    [[CSIZE:%.+]] = sext i32 [[CSZSIZE]] to i64
 
-// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 5, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SZ7:%[^,]+]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* [[MAPT7]], i32 0, i32 0), i8** null, i8** null)
+// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 {{.+}}, i32 {{.+}}, i32 {{.+}}, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CHECK-DAG:   [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CHECK-DAG:   store i8** [[BPR:%.+]], i8*** [[BPARG]]
+// CHECK-DAG:   [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CHECK-DAG:   store i8** [[PR:%.+]], i8*** [[PARG]]
+// CHECK-DAG:   [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CHECK-DAG:   store i64* [[SZ7:%.+]], i64** [[SARG]]
 // CHECK-DAG:   [[BPR]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[BP:%.+]], i32 0, i32 0
 // CHECK-DAG:   [[PR]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[P:%.+]], i32 0, i32 0
 // CHECK-DAG:   [[SZ7]] = getelementptr inbounds [5 x i64], [5 x i64]* [[PSZ:%.+]], i32 0, i32 0
@@ -631,7 +657,11 @@ int bar(int n){
 // CHECK:       [[IF:%.+]] = icmp sgt i32 {{[^,]+}}, 50
 // CHECK:       br i1 [[IF]], label %[[IFTHEN:[^,]+]], label %[[IFELSE:[^,]+]]
 // CHECK:       [[IFTHEN]]
-// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 4, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* [[SIZET6]], i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* [[MAPT6]], i32 0, i32 0), i8** null, i8** null)
+// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 {{.+}}, i32 {{.+}}, i32 {{.+}}, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CHECK-DAG:   [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CHECK-DAG:   store i8** [[BPR:%.+]], i8*** [[BPARG]]
+// CHECK-DAG:   [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CHECK-DAG:   store i8** [[PR:%.+]], i8*** [[PARG]]
 // CHECK-DAG:   [[BPR]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[BP:%.+]], i32 0, i32 0
 // CHECK-DAG:   [[PR]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[P:%.+]], i32 0, i32 0
 
@@ -682,7 +712,11 @@ int bar(int n){
 // CHECK:       [[IF:%.+]] = icmp sgt i32 {{[^,]+}}, 40
 // CHECK:       br i1 [[IF]], label %[[IFTHEN:[^,]+]], label %[[IFELSE:[^,]+]]
 // CHECK:       [[IFTHEN]]
-// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 3, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* [[SIZET5]], i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* [[MAPT5]], i32 0, i32 0), i8** null, i8** null)
+// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 {{.+}}, i32 {{.+}}, i32 {{.+}}, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CHECK-DAG:   [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CHECK-DAG:   store i8** [[BPR:%.+]], i8*** [[BPARG]]
+// CHECK-DAG:   [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CHECK-DAG:   store i8** [[PR:%.+]], i8*** [[PARG]]
 // CHECK-DAG:   [[BPR]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[BP:%.+]], i32 0, i32 0
 // CHECK-DAG:   [[PR]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[P:%.+]], i32 0, i32 0
 
@@ -742,7 +776,7 @@ int bar(int n){
 
 // OMP45:       [[BPR:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[BP]], i32 0, i32 0
 // OMP45:       [[PR:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[P]], i32 0, i32 0
-// OMP45:       [[RET:%.+]] = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 1, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZET9]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPT10]], i32 0, i32 0), i8** null, i8** null)
+// OMP45:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 {{.+}}, i32 {{.+}}, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
 // OMP45-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // OMP45-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // OMP45:       [[FAIL]]
@@ -839,7 +873,7 @@ int bar(int n){
 
 // OMP50:       [[BPR:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[BP]], i32 0, i32 0
 // OMP50:       [[PR:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[P]], i32 0, i32 0
-// OMP50:       [[RET:%.+]] = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 1, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZET9]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPT10]], i32 0, i32 0), i8** null, i8** null)
+// OMP50:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 {{.+}}, i32 {{.+}}, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
 // OMP50-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // OMP50-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // OMP50:       [[FAIL]]

diff --git a/clang/test/OpenMP/target_codegen_global_capture.cpp b/clang/test/OpenMP/target_codegen_global_capture.cpp
index ce2e7ec..cbf564e 100644 (file)
--- a/clang/test/OpenMP/target_codegen_global_capture.cpp
+++ b/clang/test/OpenMP/target_codegen_global_capture.cpp
@@ -313,9 +313,26 @@ int tbar2(short a, short b, short c, short d){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP65]], align 8
 // CHECK1-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP68:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3foossss_l49.region_id, i32 9, i8** [[TMP66]], i8** [[TMP67]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null)
-// CHECK1-NEXT:    [[TMP69:%.*]] = icmp ne i32 [[TMP68]], 0
-// CHECK1-NEXT:    br i1 [[TMP69]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK1-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 9, i32* [[TMP69]], align 4
+// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP66]], i8*** [[TMP70]], align 8
+// CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP67]], i8*** [[TMP71]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP73]], align 8
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP75]], align 8
+// CHECK1-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3foossss_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK1-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3foossss_l49(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], i64 [[TMP7]], i64 [[TMP9]], i64 [[TMP11]], i64 [[TMP13]], i64 [[TMP15]], i64 [[TMP17]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -325,28 +342,28 @@ int tbar2(short a, short b, short c, short d){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3foossss_l49(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], i64 [[TMP7]], i64 [[TMP9]], i64 [[TMP11]], i64 [[TMP13]], i64 [[TMP15]], i64 [[TMP17]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP70:%.*]] = load i16, i16* [[A_ADDR]], align 2
-// CHECK1-NEXT:    [[CONV14:%.*]] = sext i16 [[TMP70]] to i32
-// CHECK1-NEXT:    [[TMP71:%.*]] = load i16, i16* [[B_ADDR]], align 2
-// CHECK1-NEXT:    [[CONV15:%.*]] = sext i16 [[TMP71]] to i32
+// CHECK1-NEXT:    [[TMP78:%.*]] = load i16, i16* [[A_ADDR]], align 2
+// CHECK1-NEXT:    [[CONV14:%.*]] = sext i16 [[TMP78]] to i32
+// CHECK1-NEXT:    [[TMP79:%.*]] = load i16, i16* [[B_ADDR]], align 2
+// CHECK1-NEXT:    [[CONV15:%.*]] = sext i16 [[TMP79]] to i32
 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV14]], [[CONV15]]
-// CHECK1-NEXT:    [[TMP72:%.*]] = load i16, i16* [[C_ADDR]], align 2
-// CHECK1-NEXT:    [[CONV16:%.*]] = sext i16 [[TMP72]] to i32
+// CHECK1-NEXT:    [[TMP80:%.*]] = load i16, i16* [[C_ADDR]], align 2
+// CHECK1-NEXT:    [[CONV16:%.*]] = sext i16 [[TMP80]] to i32
 // CHECK1-NEXT:    [[ADD17:%.*]] = add nsw i32 [[ADD]], [[CONV16]]
-// CHECK1-NEXT:    [[TMP73:%.*]] = load i16, i16* [[D_ADDR]], align 2
-// CHECK1-NEXT:    [[CONV18:%.*]] = sext i16 [[TMP73]] to i32
+// CHECK1-NEXT:    [[TMP81:%.*]] = load i16, i16* [[D_ADDR]], align 2
+// CHECK1-NEXT:    [[CONV18:%.*]] = sext i16 [[TMP81]] to i32
 // CHECK1-NEXT:    [[ADD19:%.*]] = add nsw i32 [[ADD17]], [[CONV18]]
-// CHECK1-NEXT:    [[TMP74:%.*]] = load float, float* @_ZZ3foossssE2Sa, align 4
-// CHECK1-NEXT:    [[CONV20:%.*]] = fptosi float [[TMP74]] to i32
+// CHECK1-NEXT:    [[TMP82:%.*]] = load float, float* @_ZZ3foossssE2Sa, align 4
+// CHECK1-NEXT:    [[CONV20:%.*]] = fptosi float [[TMP82]] to i32
 // CHECK1-NEXT:    [[ADD21:%.*]] = add nsw i32 [[ADD19]], [[CONV20]]
-// CHECK1-NEXT:    [[TMP75:%.*]] = load float, float* @_ZZ3foossssE2Sb, align 4
-// CHECK1-NEXT:    [[CONV22:%.*]] = fptosi float [[TMP75]] to i32
+// CHECK1-NEXT:    [[TMP83:%.*]] = load float, float* @_ZZ3foossssE2Sb, align 4
+// CHECK1-NEXT:    [[CONV22:%.*]] = fptosi float [[TMP83]] to i32
 // CHECK1-NEXT:    [[ADD23:%.*]] = add nsw i32 [[ADD21]], [[CONV22]]
-// CHECK1-NEXT:    [[TMP76:%.*]] = load float, float* @_ZZ3foossssE2Sc, align 4
-// CHECK1-NEXT:    [[CONV24:%.*]] = fptosi float [[TMP76]] to i32
+// CHECK1-NEXT:    [[TMP84:%.*]] = load float, float* @_ZZ3foossssE2Sc, align 4
+// CHECK1-NEXT:    [[CONV24:%.*]] = fptosi float [[TMP84]] to i32
 // CHECK1-NEXT:    [[ADD25:%.*]] = add nsw i32 [[ADD23]], [[CONV24]]
-// CHECK1-NEXT:    [[TMP77:%.*]] = load float, float* @_ZZ3foossssE2Sd, align 4
-// CHECK1-NEXT:    [[CONV26:%.*]] = fptosi float [[TMP77]] to i32
+// CHECK1-NEXT:    [[TMP85:%.*]] = load float, float* @_ZZ3foossssE2Sd, align 4
+// CHECK1-NEXT:    [[CONV26:%.*]] = fptosi float [[TMP85]] to i32
 // CHECK1-NEXT:    [[ADD27:%.*]] = add nsw i32 [[ADD25]], [[CONV26]]
 // CHECK1-NEXT:    ret i32 [[ADD27]]
 //
@@ -649,9 +666,26 @@ int tbar2(short a, short b, short c, short d){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP69]], align 8
 // CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP72:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3barssss_l94.region_id, i32 9, i8** [[TMP70]], i8** [[TMP71]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_sizes.3, i32 0, i32 0), i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null)
-// CHECK1-NEXT:    [[TMP73:%.*]] = icmp ne i32 [[TMP72]], 0
-// CHECK1-NEXT:    br i1 [[TMP73]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP72]], align 4
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 9, i32* [[TMP73]], align 4
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP70]], i8*** [[TMP74]], align 8
+// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP71]], i8*** [[TMP75]], align 8
+// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_sizes.3, i32 0, i32 0), i64** [[TMP76]], align 8
+// CHECK1-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP77]], align 8
+// CHECK1-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP78]], align 8
+// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP79]], align 8
+// CHECK1-NEXT:    [[TMP80:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3barssss_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP81:%.*]] = icmp ne i32 [[TMP80]], 0
+// CHECK1-NEXT:    br i1 [[TMP81]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3barssss_l94(i64 [[TMP5]], i64 [[TMP7]], i64 [[TMP9]], i64 [[TMP11]], i64 [[TMP13]], i64 [[TMP15]], i64 [[TMP17]], i64 [[TMP19]], i64 [[TMP21]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -981,9 +1015,26 @@ int tbar2(short a, short b, short c, short d){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP69]], align 8
 // CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP72:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z4tbarIsEiT_S0_S0_S0__l145.region_id, i32 9, i8** [[TMP70]], i8** [[TMP71]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null)
-// CHECK1-NEXT:    [[TMP73:%.*]] = icmp ne i32 [[TMP72]], 0
-// CHECK1-NEXT:    br i1 [[TMP73]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP72]], align 4
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 9, i32* [[TMP73]], align 4
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP70]], i8*** [[TMP74]], align 8
+// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP71]], i8*** [[TMP75]], align 8
+// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP76]], align 8
+// CHECK1-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP77]], align 8
+// CHECK1-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP78]], align 8
+// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP79]], align 8
+// CHECK1-NEXT:    [[TMP80:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z4tbarIsEiT_S0_S0_S0__l145.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP81:%.*]] = icmp ne i32 [[TMP80]], 0
+// CHECK1-NEXT:    br i1 [[TMP81]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z4tbarIsEiT_S0_S0_S0__l145(i64 [[TMP5]], i64 [[TMP7]], i64 [[TMP9]], i64 [[TMP11]], i64 [[TMP13]], i64 [[TMP15]], i64 [[TMP17]], i64 [[TMP19]], i64 [[TMP21]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1240,9 +1291,26 @@ int tbar2(short a, short b, short c, short d){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP59]], align 4
 // CHECK3-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP62:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3foossss_l49.region_id, i32 9, i8** [[TMP60]], i8** [[TMP61]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null)
-// CHECK3-NEXT:    [[TMP63:%.*]] = icmp ne i32 [[TMP62]], 0
-// CHECK3-NEXT:    br i1 [[TMP63]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP62]], align 4
+// CHECK3-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 9, i32* [[TMP63]], align 4
+// CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP60]], i8*** [[TMP64]], align 4
+// CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP61]], i8*** [[TMP65]], align 4
+// CHECK3-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP66]], align 4
+// CHECK3-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP67]], align 4
+// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP68]], align 4
+// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3foossss_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP71:%.*]] = icmp ne i32 [[TMP70]], 0
+// CHECK3-NEXT:    br i1 [[TMP71]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3foossss_l49(i32 [[TMP1]], double* @Gb, i32 [[TMP3]], double* @Gc, i32 [[TMP5]], i32 [[TMP7]], i32 [[TMP9]], double* @Gd, i32 [[TMP11]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1252,28 +1320,28 @@ int tbar2(short a, short b, short c, short d){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3foossss_l49(i32 [[TMP1]], double* @Gb, i32 [[TMP3]], double* @Gc, i32 [[TMP5]], i32 [[TMP7]], i32 [[TMP9]], double* @Gd, i32 [[TMP11]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP64:%.*]] = load i16, i16* [[A_ADDR]], align 2
-// CHECK3-NEXT:    [[CONV11:%.*]] = sext i16 [[TMP64]] to i32
-// CHECK3-NEXT:    [[TMP65:%.*]] = load i16, i16* [[B_ADDR]], align 2
-// CHECK3-NEXT:    [[CONV12:%.*]] = sext i16 [[TMP65]] to i32
+// CHECK3-NEXT:    [[TMP72:%.*]] = load i16, i16* [[A_ADDR]], align 2
+// CHECK3-NEXT:    [[CONV11:%.*]] = sext i16 [[TMP72]] to i32
+// CHECK3-NEXT:    [[TMP73:%.*]] = load i16, i16* [[B_ADDR]], align 2
+// CHECK3-NEXT:    [[CONV12:%.*]] = sext i16 [[TMP73]] to i32
 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV11]], [[CONV12]]
-// CHECK3-NEXT:    [[TMP66:%.*]] = load i16, i16* [[C_ADDR]], align 2
-// CHECK3-NEXT:    [[CONV13:%.*]] = sext i16 [[TMP66]] to i32
+// CHECK3-NEXT:    [[TMP74:%.*]] = load i16, i16* [[C_ADDR]], align 2
+// CHECK3-NEXT:    [[CONV13:%.*]] = sext i16 [[TMP74]] to i32
 // CHECK3-NEXT:    [[ADD14:%.*]] = add nsw i32 [[ADD]], [[CONV13]]
-// CHECK3-NEXT:    [[TMP67:%.*]] = load i16, i16* [[D_ADDR]], align 2
-// CHECK3-NEXT:    [[CONV15:%.*]] = sext i16 [[TMP67]] to i32
+// CHECK3-NEXT:    [[TMP75:%.*]] = load i16, i16* [[D_ADDR]], align 2
+// CHECK3-NEXT:    [[CONV15:%.*]] = sext i16 [[TMP75]] to i32
 // CHECK3-NEXT:    [[ADD16:%.*]] = add nsw i32 [[ADD14]], [[CONV15]]
-// CHECK3-NEXT:    [[TMP68:%.*]] = load float, float* @_ZZ3foossssE2Sa, align 4
-// CHECK3-NEXT:    [[CONV17:%.*]] = fptosi float [[TMP68]] to i32
+// CHECK3-NEXT:    [[TMP76:%.*]] = load float, float* @_ZZ3foossssE2Sa, align 4
+// CHECK3-NEXT:    [[CONV17:%.*]] = fptosi float [[TMP76]] to i32
 // CHECK3-NEXT:    [[ADD18:%.*]] = add nsw i32 [[ADD16]], [[CONV17]]
-// CHECK3-NEXT:    [[TMP69:%.*]] = load float, float* @_ZZ3foossssE2Sb, align 4
-// CHECK3-NEXT:    [[CONV19:%.*]] = fptosi float [[TMP69]] to i32
+// CHECK3-NEXT:    [[TMP77:%.*]] = load float, float* @_ZZ3foossssE2Sb, align 4
+// CHECK3-NEXT:    [[CONV19:%.*]] = fptosi float [[TMP77]] to i32
 // CHECK3-NEXT:    [[ADD20:%.*]] = add nsw i32 [[ADD18]], [[CONV19]]
-// CHECK3-NEXT:    [[TMP70:%.*]] = load float, float* @_ZZ3foossssE2Sc, align 4
-// CHECK3-NEXT:    [[CONV21:%.*]] = fptosi float [[TMP70]] to i32
+// CHECK3-NEXT:    [[TMP78:%.*]] = load float, float* @_ZZ3foossssE2Sc, align 4
+// CHECK3-NEXT:    [[CONV21:%.*]] = fptosi float [[TMP78]] to i32
 // CHECK3-NEXT:    [[ADD22:%.*]] = add nsw i32 [[ADD20]], [[CONV21]]
-// CHECK3-NEXT:    [[TMP71:%.*]] = load float, float* @_ZZ3foossssE2Sd, align 4
-// CHECK3-NEXT:    [[CONV23:%.*]] = fptosi float [[TMP71]] to i32
+// CHECK3-NEXT:    [[TMP79:%.*]] = load float, float* @_ZZ3foossssE2Sd, align 4
+// CHECK3-NEXT:    [[CONV23:%.*]] = fptosi float [[TMP79]] to i32
 // CHECK3-NEXT:    [[ADD24:%.*]] = add nsw i32 [[ADD22]], [[CONV23]]
 // CHECK3-NEXT:    ret i32 [[ADD24]]
 //
@@ -1570,9 +1638,26 @@ int tbar2(short a, short b, short c, short d){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP63]], align 4
 // CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3barssss_l94.region_id, i32 9, i8** [[TMP64]], i8** [[TMP65]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_sizes.3, i32 0, i32 0), i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null)
-// CHECK3-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
-// CHECK3-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK3-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 9, i32* [[TMP67]], align 4
+// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_sizes.3, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK3-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP71]], align 4
+// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP73]], align 4
+// CHECK3-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3barssss_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK3-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3barssss_l94(i32 [[TMP5]], double* @Gb, i32 [[TMP7]], double* @Gc, i32 [[TMP9]], i32 [[TMP11]], i32 [[TMP13]], double* @Gd, i32 [[TMP15]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1896,9 +1981,26 @@ int tbar2(short a, short b, short c, short d){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP63]], align 4
 // CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z4tbarIsEiT_S0_S0_S0__l145.region_id, i32 9, i8** [[TMP64]], i8** [[TMP65]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null)
-// CHECK3-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
-// CHECK3-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK3-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 9, i32* [[TMP67]], align 4
+// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK3-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP71]], align 4
+// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP73]], align 4
+// CHECK3-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z4tbarIsEiT_S0_S0_S0__l145.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK3-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z4tbarIsEiT_S0_S0_S0__l145(i32 [[TMP5]], double* @Gb, i32 [[TMP7]], double* @Gc, i32 [[TMP9]], i32 [[TMP11]], i32 [[TMP13]], double* @Gd, i32 [[TMP15]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/target_data_member_codegen.cpp b/clang/test/OpenMP/target_data_member_codegen.cpp
index b0e42de..b735271 100644 (file)
--- a/clang/test/OpenMP/target_data_member_codegen.cpp
+++ b/clang/test/OpenMP/target_data_member_codegen.cpp
@@ -33,7 +33,10 @@ struct S {
 
 // CHECK-LABEL: @main
 int main() {
-  // CHECK: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @.{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZES]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPTYPES]], i32 0, i32 0), i8** null, i8** null)
+    // CHECK: store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0)
+    // CHECK: store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0)
+    // CHECK: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
+
     S s;
 #pragma omp target map(s.p)
     { }

diff --git a/clang/test/OpenMP/target_data_use_device_ptr_if_codegen.cpp b/clang/test/OpenMP/target_data_use_device_ptr_if_codegen.cpp
index 3956c46..0f35ed3 100644 (file)
--- a/clang/test/OpenMP/target_data_use_device_ptr_if_codegen.cpp
+++ b/clang/test/OpenMP/target_data_use_device_ptr_if_codegen.cpp
@@ -32,8 +32,8 @@ void add_one(float *b, int dm)
   // CK1-NOT: store float* [[VAL]], float** {{%.+}},
   // CK1:     store float* [[VAL]], float** [[PVT:%.+]],
   // CK1:     [[TT:%.+]] = load float*, float** [[PVT]],
-  // CK1:     call i32 @__tgt_target{{.+}}[[MTYPE01]]
-  // CK1:     call i32 @__tgt_target{{.+}}[[MTYPE02]]
+  // CK1:     call i32 @__tgt_target{{.+}}
+  // CK1:     call i32 @__tgt_target{{.+}}
   // CK1:     call void @__tgt_target_data_end{{.+}}[[MTYPE00]]
 #pragma omp target data map(tofrom:b[:1]) use_device_ptr(b) if(dm == 0)
   {

diff --git a/clang/test/OpenMP/target_defaultmap_codegen_01.cpp b/clang/test/OpenMP/target_defaultmap_codegen_01.cpp
index 5000679..9269f9a 100644 (file)
--- a/clang/test/OpenMP/target_defaultmap_codegen_01.cpp
+++ b/clang/test/OpenMP/target_defaultmap_codegen_01.cpp
@@ -30,18 +30,23 @@
 void implicit_maps_double_complex (int a){
   double _Complex dc = (double)a;
 
-  // CK1-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { double, double }**
-  // CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { double, double }**
-  // CK1-DAG: store { double, double }* [[PTR:%[^,]+]], { double, double }** [[CBP1]]
-  // CK1-DAG: store { double, double }* [[PTR]], { double, double }** [[CP1]]
-
-  // CK1: call void [[KERNEL:@.+]]({ double, double }* [[PTR]])
-  #pragma omp target defaultmap(alloc:scalar)
+// CK1-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK1-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK1-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK1-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK1-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { double, double }**
+// CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { double, double }**
+// CK1-DAG: store { double, double }* [[PTR:%[^,]+]], { double, double }** [[CBP1]]
+// CK1-DAG: store { double, double }* [[PTR]], { double, double }** [[CP1]]
+
+// CK1: call void [[KERNEL:@.+]]({ double, double }* [[PTR]])
+#pragma omp target defaultmap(alloc \
+                              : scalar)
   {
    dc *= dc;
   }
@@ -80,18 +85,23 @@ void implicit_maps_double_complex (int a){
 void implicit_maps_double_complex (int a){
   double _Complex dc = (double)a;
 
-  // CK2-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK2-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK2-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK2-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK2-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK2-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { double, double }**
-  // CK2-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { double, double }**
-  // CK2-DAG: store { double, double }* [[PTR:%[^,]+]], { double, double }** [[CBP1]]
-  // CK2-DAG: store { double, double }* [[PTR]], { double, double }** [[CP1]]
-
-  // CK2: call void [[KERNEL:@.+]]({ double, double }* [[PTR]])
-  #pragma omp target defaultmap(to:scalar)
+// CK2-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK2-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK2-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK2-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK2-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK2-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK2-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK2-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK2-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK2-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { double, double }**
+// CK2-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { double, double }**
+// CK2-DAG: store { double, double }* [[PTR:%[^,]+]], { double, double }** [[CBP1]]
+// CK2-DAG: store { double, double }* [[PTR]], { double, double }** [[CP1]]
+
+// CK2: call void [[KERNEL:@.+]]({ double, double }* [[PTR]])
+#pragma omp target defaultmap(to \
+                              : scalar)
   {
    dc *= dc;
   }
@@ -130,18 +140,23 @@ void implicit_maps_double_complex (int a){
 void implicit_maps_double_complex (int a){
   double _Complex dc = (double)a;
 
-  // CK3-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK3-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK3-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK3-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK3-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK3-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { double, double }**
-  // CK3-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { double, double }**
-  // CK3-DAG: store { double, double }* [[PTR:%[^,]+]], { double, double }** [[CBP1]]
-  // CK3-DAG: store { double, double }* [[PTR]], { double, double }** [[CP1]]
-
-  // CK3: call void [[KERNEL:@.+]]({ double, double }* [[PTR]])
-  #pragma omp target defaultmap(from:scalar)
+// CK3-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK3-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK3-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK3-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK3-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK3-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK3-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK3-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK3-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK3-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { double, double }**
+// CK3-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { double, double }**
+// CK3-DAG: store { double, double }* [[PTR:%[^,]+]], { double, double }** [[CBP1]]
+// CK3-DAG: store { double, double }* [[PTR]], { double, double }** [[CP1]]
+
+// CK3: call void [[KERNEL:@.+]]({ double, double }* [[PTR]])
+#pragma omp target defaultmap(from \
+                              : scalar)
   {
    dc *= dc;
   }
@@ -185,28 +200,33 @@ void implicit_maps_double_complex (int a){
 void implicit_maps_double (int a){
   double d = (double)a;
 
-  // CK4-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK4-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK4-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK4-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK4-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-
-  // CK4-64-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
-  // CK4-64-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
-  // CK4-64-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
-  // CK4-64-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
-  // CK4-64-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
-  // CK4-64-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to double*
-  // CK4-64-64-DAG: store double {{.+}}, double* [[CADDR]],
-
-  // CK4-32-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
-  // CK4-32-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
-  // CK4-32-DAG: store double* [[DECL:%[^,]+]], double** [[CBP1]]
-  // CK4-32-DAG: store double* [[DECL]], double** [[CP1]]
-
-  // CK4-64: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
-  // CK4-32: call void [[KERNEL:@.+]](double* [[DECL]])
-  #pragma omp target defaultmap(firstprivate:scalar)
+// CK4-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK4-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK4-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK4-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK4-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK4-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK4-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK4-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK4-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+
+// CK4-64-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
+// CK4-64-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
+// CK4-64-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
+// CK4-64-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
+// CK4-64-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
+// CK4-64-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to double*
+// CK4-64-64-DAG: store double {{.+}}, double* [[CADDR]],
+
+// CK4-32-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
+// CK4-32-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
+// CK4-32-DAG: store double* [[DECL:%[^,]+]], double** [[CBP1]]
+// CK4-32-DAG: store double* [[DECL]], double** [[CP1]]
+
+// CK4-64: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
+// CK4-32: call void [[KERNEL:@.+]](double* [[DECL]])
+#pragma omp target defaultmap(firstprivate \
+                              : scalar)
   {
     d += 1.0;
   }
@@ -252,18 +272,23 @@ void implicit_maps_double (int a){
 void implicit_maps_array (int a){
   double darr[2] = {(double)a, (double)a};
 
-  // CK5-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK5-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK5-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK5-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK5-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK5-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [2 x double]**
-  // CK5-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [2 x double]**
-  // CK5-DAG: store [2 x double]* [[DECL:%[^,]+]], [2 x double]** [[CBP1]]
-  // CK5-DAG: store [2 x double]* [[DECL]], [2 x double]** [[CP1]]
-
-  // CK5: call void [[KERNEL:@.+]]([2 x double]* [[DECL]])
-  #pragma omp target defaultmap(alloc: aggregate)
+// CK5-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK5-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK5-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK5-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK5-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK5-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK5-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK5-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK5-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK5-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [2 x double]**
+// CK5-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [2 x double]**
+// CK5-DAG: store [2 x double]* [[DECL:%[^,]+]], [2 x double]** [[CBP1]]
+// CK5-DAG: store [2 x double]* [[DECL]], [2 x double]** [[CP1]]
+
+// CK5: call void [[KERNEL:@.+]]([2 x double]* [[DECL]])
+#pragma omp target defaultmap(alloc \
+                              : aggregate)
   {
     darr[0] += 1.0;
     darr[1] += 1.0;
@@ -303,7 +328,11 @@ void implicit_maps_array (int a){
 void implicit_maps_array (int a){
   double darr[2] = {(double)a, (double)a};
 
-  // CK6-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
+  // CK6-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK6-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK6-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK6-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK6-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
   // CK6-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
   // CK6-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
   // CK6-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
@@ -354,18 +383,23 @@ void implicit_maps_array (int a){
 void implicit_maps_array (int a){
   double darr[2] = {(double)a, (double)a};
 
-  // CK7-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK7-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK7-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK7-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK7-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK7-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [2 x double]**
-  // CK7-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [2 x double]**
-  // CK7-DAG: store [2 x double]* [[DECL:%[^,]+]], [2 x double]** [[CBP1]]
-  // CK7-DAG: store [2 x double]* [[DECL]], [2 x double]** [[CP1]]
-
-  // CK7: call void [[KERNEL:@.+]]([2 x double]* [[DECL]])
-  #pragma omp target defaultmap(from: aggregate)
+// CK7-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK7-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK7-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK7-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK7-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK7-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK7-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK7-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK7-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK7-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [2 x double]**
+// CK7-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [2 x double]**
+// CK7-DAG: store [2 x double]* [[DECL:%[^,]+]], [2 x double]** [[CBP1]]
+// CK7-DAG: store [2 x double]* [[DECL]], [2 x double]** [[CP1]]
+
+// CK7: call void [[KERNEL:@.+]]([2 x double]* [[DECL]])
+#pragma omp target defaultmap(from \
+                              : aggregate)
   {
     darr[0] += 1.0;
     darr[1] += 1.0;
@@ -405,7 +439,11 @@ void implicit_maps_array (int a){
 void implicit_maps_array (int a){
   double darr[2] = {(double)a, (double)a};
 
-  // CK8-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
+  // CK8-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK8-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK8-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK8-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK8-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
   // CK8-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
   // CK8-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
   // CK8-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
@@ -458,20 +496,25 @@ void implicit_maps_array (int a){
 void zero_size_section_and_private_maps (int ii){
   int pvtArr[10];
 
-  // Region 09
-  // CK9-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE09]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE09]]{{.+}}, i8** null, i8** null)
-  // CK9-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK9-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK9-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK9-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK9-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [10 x i32]**
-  // CK9-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [10 x i32]**
-  // CK9-DAG: store [10 x i32]* [[VAR0:%.+]], [10 x i32]** [[CBP0]]
-  // CK9-DAG: store [10 x i32]* [[VAR0]], [10 x i32]** [[CP0]]
-
-  // CK9: call void [[CALL09:@.+]]([10 x i32]* {{[^,]+}})
-  #pragma omp target defaultmap(firstprivate:aggregate)
+// Region 09
+// CK9-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK9-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK9-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK9-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK9-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK9-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK9-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK9-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK9-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK9-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [10 x i32]**
+// CK9-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [10 x i32]**
+// CK9-DAG: store [10 x i32]* [[VAR0:%.+]], [10 x i32]** [[CBP0]]
+// CK9-DAG: store [10 x i32]* [[VAR0]], [10 x i32]** [[CP0]]
+
+// CK9: call void [[CALL09:@.+]]([10 x i32]* {{[^,]+}})
+#pragma omp target defaultmap(firstprivate \
+                              : aggregate)
   {
     pvtArr[5]++;
   }
@@ -508,19 +551,24 @@ void zero_size_section_and_private_maps (int ii){
 void explicit_maps_single (){
   int *pa;
 
-  // CK10-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE]]{{.+}}, i8** null, i8** null)
-  // CK10-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK10-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK10-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK10-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK10-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
-  // CK10-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
-  // CK10-DAG: store i32** [[VAR0:%.+]], i32*** [[CBP0]]
-  // CK10-DAG: store i32** [[VAR0]], i32*** [[CP0]]
-
-  // CK10: call void [[CALL:@.+]](i32** {{[^,]+}})
-  #pragma omp target defaultmap(alloc: pointer)
+// CK10-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK10-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK10-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK10-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK10-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK10-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK10-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK10-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK10-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK10-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
+// CK10-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
+// CK10-DAG: store i32** [[VAR0:%.+]], i32*** [[CBP0]]
+// CK10-DAG: store i32** [[VAR0]], i32*** [[CP0]]
+
+// CK10: call void [[CALL:@.+]](i32** {{[^,]+}})
+#pragma omp target defaultmap(alloc \
+                              : pointer)
   {
     pa[50]++;
   }
@@ -556,19 +604,24 @@ void explicit_maps_single (){
 void explicit_maps_single (){
   int *pa;
 
-  // CK11-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE09]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE09]]{{.+}}, i8** null, i8** null)
-  // CK11-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK11-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK11-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK11-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK11-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
-  // CK11-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
-  // CK11-DAG: store i32** [[VAR0:%.+]], i32*** [[CBP0]]
-  // CK11-DAG: store i32** [[VAR0]], i32*** [[CP0]]
-
-  // CK11: call void [[CALL09:@.+]](i32** {{[^,]+}})
-  #pragma omp target defaultmap(to: pointer)
+// CK11-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK11-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK11-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK11-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK11-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK11-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK11-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK11-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK11-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK11-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
+// CK11-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
+// CK11-DAG: store i32** [[VAR0:%.+]], i32*** [[CBP0]]
+// CK11-DAG: store i32** [[VAR0]], i32*** [[CP0]]
+
+// CK11: call void [[CALL09:@.+]](i32** {{[^,]+}})
+#pragma omp target defaultmap(to \
+                              : pointer)
   {
     pa[50]++;
   }
@@ -604,19 +657,24 @@ void explicit_maps_single (){
 void explicit_maps_single (){
   int *pa;
 
-  // CK12-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE09]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE09]]{{.+}}, i8** null, i8** null)
-  // CK12-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK12-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK12-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK12-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK12-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
-  // CK12-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
-  // CK12-DAG: store i32** [[VAR0:%.+]], i32*** [[CBP0]]
-  // CK12-DAG: store i32** [[VAR0]], i32*** [[CP0]]
-
-  // CK12: call void [[CALL09:@.+]](i32** {{[^,]+}})
-  #pragma omp target defaultmap(from: pointer)
+// CK12-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK12-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK12-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK12-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK12-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK12-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK12-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK12-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK12-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK12-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
+// CK12-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
+// CK12-DAG: store i32** [[VAR0:%.+]], i32*** [[CBP0]]
+// CK12-DAG: store i32** [[VAR0]], i32*** [[CP0]]
+
+// CK12: call void [[CALL09:@.+]](i32** {{[^,]+}})
+#pragma omp target defaultmap(from \
+                              : pointer)
   {
     pa[50]++;
   }
@@ -652,19 +710,24 @@ void explicit_maps_single (){
 void explicit_maps_single (){
   int *pa;
 
-  // CK13-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE09]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE09]]{{.+}}, i8** null, i8** null)
-  // CK13-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK13-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK13-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK13-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK13-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
-  // CK13-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
-  // CK13-DAG: store i32** [[VAR0:%.+]], i32*** [[CBP0]]
-  // CK13-DAG: store i32** [[VAR0]], i32*** [[CP0]]
-
-  // CK13: call void [[CALL09:@.+]](i32** {{[^,]+}})
-  #pragma omp target defaultmap(tofrom: pointer)
+// CK13-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK13-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK13-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK13-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK13-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK13-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK13-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK13-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK13-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK13-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
+// CK13-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
+// CK13-DAG: store i32** [[VAR0:%.+]], i32*** [[CBP0]]
+// CK13-DAG: store i32** [[VAR0]], i32*** [[CP0]]
+
+// CK13: call void [[CALL09:@.+]](i32** {{[^,]+}})
+#pragma omp target defaultmap(tofrom \
+                              : pointer)
   {
     pa[50]++;
   }
@@ -700,19 +763,24 @@ void explicit_maps_single (){
 void explicit_maps_single (){
   int *pa;
 
-  // CK14-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE09]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE09]]{{.+}}, i8** null, i8** null)
-  // CK14-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK14-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK14-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK14-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK14-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK14-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK14-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
-  // CK14-DAG: store i32* [[VAR0]], i32** [[CP0]]
-
-  // CK14: call void [[CALL09:@.+]](i32* {{[^,]+}})
-  #pragma omp target defaultmap(firstprivate: pointer)
+// CK14-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK14-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK14-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK14-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK14-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK14-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK14-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK14-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK14-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK14-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK14-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK14-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
+// CK14-DAG: store i32* [[VAR0]], i32** [[CP0]]
+
+// CK14: call void [[CALL09:@.+]](i32* {{[^,]+}})
+#pragma omp target defaultmap(firstprivate \
+                              : pointer)
   {
     pa[50]++;
   }
@@ -750,37 +818,44 @@ void explicit_maps_single (){
 void implicit_maps_variable_length_array (int a){
   double vla[2][a];
 
-  // CK15-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 3, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], i64* [[SGEP:%[^,]+]], {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK15-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK15-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK15-DAG: [[SGEP]] = getelementptr inbounds {{.+}}[[SS:%[^,]+]], i32 0, i32 0
-
-  // CK15-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK15-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK15-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[sz:64|32]]*
-  // CK15-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[sz]]*
-  // CK15-DAG: store i[[sz]] 2, i[[sz]]* [[CBP0]]
-  // CK15-DAG: store i[[sz]] 2, i[[sz]]* [[CP0]]
-
-  // CK15-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
-  // CK15-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
-  // CK15-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
-  // CK15-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
-  // CK15-DAG: store i[[sz]] [[VAL:%.+]], i[[sz]]* [[CBP1]]
-  // CK15-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
-
-  // CK15-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 2
-  // CK15-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 2
-  // CK15-DAG: [[S2:%.+]] = getelementptr inbounds {{.+}}[[SS]], i32 0, i32 2
-  // CK15-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to double**
-  // CK15-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double**
-  // CK15-DAG: store double* [[DECL:%.+]], double** [[CBP2]]
-  // CK15-DAG: store double* [[DECL]], double** [[CP2]]
-  // CK15-DAG: store i64 [[VALS2:%.+]], i64* [[S2]],
-  // CK15-DAG: [[VALS2]] = {{mul nuw i64 %.+, 8|sext i32 %.+ to i64}}
-
-  // CK15: call void [[KERNEL:@.+]](i[[sz]] {{.+}}, i[[sz]] {{.+}}, double* [[DECL]])
-  #pragma omp target defaultmap(alloc: aggregate)
+// CK15-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK15-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK15-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK15-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK15-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK15-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK15-DAG: store i64* [[SGEP:%.+]], i64** [[SARG]]
+// CK15-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK15-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK15-DAG: [[SGEP]] = getelementptr inbounds {{.+}}[[SS:%[^,]+]], i32 0, i32 0
+
+// CK15-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK15-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK15-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[sz:64|32]]*
+// CK15-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[sz]]*
+// CK15-DAG: store i[[sz]] 2, i[[sz]]* [[CBP0]]
+// CK15-DAG: store i[[sz]] 2, i[[sz]]* [[CP0]]
+
+// CK15-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
+// CK15-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
+// CK15-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
+// CK15-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
+// CK15-DAG: store i[[sz]] [[VAL:%.+]], i[[sz]]* [[CBP1]]
+// CK15-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
+
+// CK15-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 2
+// CK15-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 2
+// CK15-DAG: [[S2:%.+]] = getelementptr inbounds {{.+}}[[SS]], i32 0, i32 2
+// CK15-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to double**
+// CK15-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double**
+// CK15-DAG: store double* [[DECL:%.+]], double** [[CBP2]]
+// CK15-DAG: store double* [[DECL]], double** [[CP2]]
+// CK15-DAG: store i64 [[VALS2:%.+]], i64* [[S2]],
+// CK15-DAG: [[VALS2]] = {{mul nuw i64 %.+, 8|sext i32 %.+ to i64}}
+
+// CK15: call void [[KERNEL:@.+]](i[[sz]] {{.+}}, i[[sz]] {{.+}}, double* [[DECL]])
+#pragma omp target defaultmap(alloc \
+                              : aggregate)
   {
     vla[1][3] += 1.0;
   }
@@ -831,18 +906,23 @@ public:
 void implicit_maps_struct (int a){
   SSS s = {a, (double)a};
 
-  // CK16-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK16-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK16-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK16-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK16-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK16-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
-  // CK16-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[ST]]**
-  // CK16-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP1]]
-  // CK16-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CP1]]
-
-  // CK16: call void [[KERNEL:@.+]]([[ST]]* [[DECL]])
-  #pragma omp target defaultmap(alloc: aggregate)
+// CK16-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK16-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK16-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK16-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK16-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK16-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK16-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK16-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK16-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK16-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
+// CK16-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[ST]]**
+// CK16-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP1]]
+// CK16-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CP1]]
+
+// CK16: call void [[KERNEL:@.+]]([[ST]]* [[DECL]])
+#pragma omp target defaultmap(alloc \
+                              : aggregate)
   {
     s.a += 1;
     s.b += 1.0;
@@ -888,18 +968,23 @@ public:
 void implicit_maps_struct (int a){
   SSS s = {a, (double)a};
 
-  // CK17-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK17-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK17-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK17-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK17-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK17-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
-  // CK17-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[ST]]**
-  // CK17-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP1]]
-  // CK17-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CP1]]
-
-  // CK17: call void [[KERNEL:@.+]]([[ST]]* [[DECL]])
-  #pragma omp target defaultmap(to: aggregate)
+// CK17-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK17-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK17-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK17-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK17-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK17-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK17-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK17-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK17-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK17-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
+// CK17-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[ST]]**
+// CK17-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP1]]
+// CK17-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CP1]]
+
+// CK17: call void [[KERNEL:@.+]]([[ST]]* [[DECL]])
+#pragma omp target defaultmap(to \
+                              : aggregate)
   {
     s.a += 1;
     s.b += 1.0;
@@ -945,18 +1030,23 @@ public:
 void implicit_maps_struct (int a){
   SSS s = {a, (double)a};
 
-  // CK18-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK18-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK18-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK18-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK18-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK18-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
-  // CK18-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[ST]]**
-  // CK18-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP1]]
-  // CK18-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CP1]]
-
-  // CK18: call void [[KERNEL:@.+]]([[ST]]* [[DECL]])
-  #pragma omp target defaultmap(from: aggregate)
+// CK18-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK18-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK18-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK18-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK18-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK18-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK18-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK18-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK18-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK18-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
+// CK18-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[ST]]**
+// CK18-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP1]]
+// CK18-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CP1]]
+
+// CK18: call void [[KERNEL:@.+]]([[ST]]* [[DECL]])
+#pragma omp target defaultmap(from \
+                              : aggregate)
   {
     s.a += 1;
     s.b += 1.0;
@@ -1002,18 +1092,23 @@ public:
 void implicit_maps_struct (int a){
   SSS s = {a, (double)a};
 
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
-  // CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[ST]]**
-  // CK19-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP1]]
-  // CK19-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CP1]]
-
-  // CK19: call void [[KERNEL:@.+]]([[ST]]* [[DECL]])
-  #pragma omp target defaultmap(tofrom: aggregate)
+// CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK19-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK19-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK19-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
+// CK19-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[ST]]**
+// CK19-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP1]]
+// CK19-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CP1]]
+
+// CK19: call void [[KERNEL:@.+]]([[ST]]* [[DECL]])
+#pragma omp target defaultmap(tofrom \
+                              : aggregate)
   {
     s.a += 1;
     s.b += 1.0;
@@ -1058,28 +1153,33 @@ void implicit_maps_struct (int a){
 void implicit_maps_double (int a){
   double d = (double)a;
 
-  // CK20-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK20-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK20-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK20-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK20-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-
-  // CK20-64-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
-  // CK20-64-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
-  // CK20-64-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
-  // CK20-64-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
-  // CK20-64-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
-  // CK20-64-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to double*
-  // CK20-64-64-DAG: store double {{.+}}, double* [[CADDR]],
-
-  // CK20-32-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
-  // CK20-32-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
-  // CK20-32-DAG: store double* [[DECL:%[^,]+]], double** [[CBP1]]
-  // CK20-32-DAG: store double* [[DECL]], double** [[CP1]]
-
-  // CK20-64: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
-  // CK20-32: call void [[KERNEL:@.+]](double* [[DECL]])
-  #pragma omp target defaultmap(default: scalar)
+// CK20-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK20-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK20-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK20-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK20-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK20-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK20-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK20-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK20-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+
+// CK20-64-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
+// CK20-64-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
+// CK20-64-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
+// CK20-64-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
+// CK20-64-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
+// CK20-64-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to double*
+// CK20-64-64-DAG: store double {{.+}}, double* [[CADDR]],
+
+// CK20-32-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
+// CK20-32-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
+// CK20-32-DAG: store double* [[DECL:%[^,]+]], double** [[CBP1]]
+// CK20-32-DAG: store double* [[DECL]], double** [[CP1]]
+
+// CK20-64: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
+// CK20-32: call void [[KERNEL:@.+]](double* [[DECL]])
+#pragma omp target defaultmap(default \
+                              : scalar)
   {
     d += 1.0;
   }
@@ -1131,18 +1231,23 @@ public:
 void implicit_maps_struct (int a){
   SSS s = {a, (double)a};
 
-  // CK21-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK21-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK21-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK21-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK21-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK21-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
-  // CK21-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[ST]]**
-  // CK21-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP1]]
-  // CK21-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CP1]]
-
-  // CK21: call void [[KERNEL:@.+]]([[ST]]* [[DECL]])
-  #pragma omp target defaultmap(default: aggregate)
+// CK21-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK21-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK21-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK21-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK21-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK21-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK21-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK21-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK21-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK21-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
+// CK21-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[ST]]**
+// CK21-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP1]]
+// CK21-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CP1]]
+
+// CK21: call void [[KERNEL:@.+]]([[ST]]* [[DECL]])
+#pragma omp target defaultmap(default \
+                              : aggregate)
   {
     s.a += 1;
     s.b += 1.0;
@@ -1182,18 +1287,23 @@ void implicit_maps_struct (int a){
 void implicit_maps_pointer (){
   double *ddyn;
 
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK22-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK22-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK22-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
-  // CK22-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
-  // CK22-DAG: store double* [[PTR:%[^,]+]], double** [[CBP1]]
-  // CK22-DAG: store double* [[PTR]], double** [[CP1]]
-
-  // CK22: call void [[KERNEL:@.+]](double* [[PTR]])
-  #pragma omp target defaultmap(default: pointer)
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK22-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK22-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK22-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
+// CK22-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
+// CK22-DAG: store double* [[PTR:%[^,]+]], double** [[CBP1]]
+// CK22-DAG: store double* [[PTR]], double** [[CP1]]
+
+// CK22: call void [[KERNEL:@.+]](double* [[PTR]])
+#pragma omp target defaultmap(default \
+                              : pointer)
   {
     ddyn[0] += 1.0;
     ddyn[1] += 1.0;
@@ -1254,134 +1364,169 @@ void foo(float *&lr, T *&tr) {
   float *l;
   T *t;
 
-  // CK23-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES00]]{{.+}}, {{.+}}[[TYPES00]]{{.+}}, i8** null, i8** null)
-  // CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
-  // CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
-  // CK23-DAG: store double* [[VAL:%.+]], double** [[CBP1]]
-  // CK23-DAG: store double* [[VAL]], double** [[CP1]]
-  // CK23-DAG: [[VAL]] = load double*, double** [[ADDR:@g]],
-
-  // CK23: call void [[KERNEL:@.+]](double* [[VAL]])
-  #pragma omp target is_device_ptr(g) defaultmap(none:pointer)
+// CK23-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK23-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK23-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK23-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK23-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
+// CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
+// CK23-DAG: store double* [[VAL:%.+]], double** [[CBP1]]
+// CK23-DAG: store double* [[VAL]], double** [[CP1]]
+// CK23-DAG: [[VAL]] = load double*, double** [[ADDR:@g]],
+
+// CK23: call void [[KERNEL:@.+]](double* [[VAL]])
+#pragma omp target is_device_ptr(g) defaultmap(none \
+                                               : pointer)
   {
     ++g;
   }
 
-  // CK23-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES01]]{{.+}}, {{.+}}[[TYPES01]]{{.+}}, i8** null, i8** null)
-  // CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
-  // CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
-  // CK23-DAG: store float* [[VAL:%.+]], float** [[CBP1]]
-  // CK23-DAG: store float* [[VAL]], float** [[CP1]]
-  // CK23-DAG: [[VAL]] = load float*, float** [[ADDR:%.+]],
-
-  // CK23: call void [[KERNEL:@.+]](float* [[VAL]])
-  #pragma omp target is_device_ptr(l) defaultmap(none:pointer)
+// CK23-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK23-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK23-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK23-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK23-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
+// CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
+// CK23-DAG: store float* [[VAL:%.+]], float** [[CBP1]]
+// CK23-DAG: store float* [[VAL]], float** [[CP1]]
+// CK23-DAG: [[VAL]] = load float*, float** [[ADDR:%.+]],
+
+// CK23: call void [[KERNEL:@.+]](float* [[VAL]])
+#pragma omp target is_device_ptr(l) defaultmap(none \
+                                               : pointer)
   {
     ++l;
   }
 
-  // CK23-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES02]]{{.+}}, {{.+}}[[TYPES02]]{{.+}}, i8** null, i8** null)
-  // CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
-  // CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
-  // CK23-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
-  // CK23-DAG: store i32* [[VAL]], i32** [[CP1]]
-  // CK23-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
-
-  // CK23: call void [[KERNEL:@.+]](i32* [[VAL]])
-  #pragma omp target is_device_ptr(t) defaultmap(none:pointer)
+// CK23-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK23-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK23-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK23-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK23-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
+// CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
+// CK23-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
+// CK23-DAG: store i32* [[VAL]], i32** [[CP1]]
+// CK23-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
+
+// CK23: call void [[KERNEL:@.+]](i32* [[VAL]])
+#pragma omp target is_device_ptr(t) defaultmap(none \
+                                               : pointer)
   {
     ++t;
   }
 
-  // CK23-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES03]]{{.+}}, {{.+}}[[TYPES03]]{{.+}}, i8** null, i8** null)
-  // CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
-  // CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
-  // CK23-DAG: store float* [[VAL:%.+]], float** [[CBP1]]
-  // CK23-DAG: store float* [[VAL]], float** [[CP1]]
-  // CK23-DAG: [[VAL]] = load float*, float** [[ADDR:%.+]],
-  // CK23-DAG: [[ADDR]] = load float**, float*** [[ADDR2:%.+]],
-
-  // CK23: call void [[KERNEL:@.+]](float* [[VAL]])
-  #pragma omp target is_device_ptr(lr) defaultmap(none:pointer)
+// CK23-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK23-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK23-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK23-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK23-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
+// CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
+// CK23-DAG: store float* [[VAL:%.+]], float** [[CBP1]]
+// CK23-DAG: store float* [[VAL]], float** [[CP1]]
+// CK23-DAG: [[VAL]] = load float*, float** [[ADDR:%.+]],
+// CK23-DAG: [[ADDR]] = load float**, float*** [[ADDR2:%.+]],
+
+// CK23: call void [[KERNEL:@.+]](float* [[VAL]])
+#pragma omp target is_device_ptr(lr) defaultmap(none \
+                                                : pointer)
   {
     ++lr;
   }
 
-  // CK23-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES04]]{{.+}}, {{.+}}[[TYPES04]]{{.+}}, i8** null, i8** null)
-  // CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
-  // CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
-  // CK23-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
-  // CK23-DAG: store i32* [[VAL]], i32** [[CP1]]
-  // CK23-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
-  // CK23-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
-
-  // CK23: call void [[KERNEL:@.+]](i32* [[VAL]])
-  #pragma omp target is_device_ptr(tr) defaultmap(none:pointer)
+// CK23-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK23-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK23-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK23-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK23-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
+// CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
+// CK23-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
+// CK23-DAG: store i32* [[VAL]], i32** [[CP1]]
+// CK23-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
+// CK23-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
+
+// CK23: call void [[KERNEL:@.+]](i32* [[VAL]])
+#pragma omp target is_device_ptr(tr) defaultmap(none \
+                                                : pointer)
   {
     ++tr;
   }
 
-  // CK23-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES05]]{{.+}}, {{.+}}[[TYPES05]]{{.+}}, i8** null, i8** null)
-  // CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
-  // CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
-  // CK23-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
-  // CK23-DAG: store i32* [[VAL]], i32** [[CP1]]
-  // CK23-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
-  // CK23-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
-
-  // CK23: call void [[KERNEL:@.+]](i32* [[VAL]])
-  #pragma omp target is_device_ptr(tr,lr) defaultmap(none:pointer)
+// CK23-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK23-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK23-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK23-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK23-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
+// CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
+// CK23-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
+// CK23-DAG: store i32* [[VAL]], i32** [[CP1]]
+// CK23-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
+// CK23-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
+
+// CK23: call void [[KERNEL:@.+]](i32* [[VAL]])
+#pragma omp target is_device_ptr(tr, lr) defaultmap(none \
+                                                    : pointer)
   {
     ++tr;
   }
 
-  // CK23-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 2, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES06]]{{.+}}, {{.+}}[[TYPES06]]{{.+}}, i8** null, i8** null)
-  // CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
-  // CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
-  // CK23-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
-  // CK23-DAG: store i32* [[VAL]], i32** [[CP1]]
-  // CK23-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
-  // CK23-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
-
-  // CK23-DAG: [[_BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
-  // CK23-DAG: [[_P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
-  // CK23-DAG: [[_CBP1:%.+]] = bitcast i8** [[_BP1]] to float**
-  // CK23-DAG: [[_CP1:%.+]] = bitcast i8** [[_P1]] to float**
-  // CK23-DAG: store float* [[_VAL:%.+]], float** [[_CBP1]]
-  // CK23-DAG: store float* [[_VAL]], float** [[_CP1]]
-  // CK23-DAG: [[_VAL]] = load float*, float** [[_ADDR:%.+]],
-  // CK23-DAG: [[_ADDR]] = load float**, float*** [[_ADDR2:%.+]],
-
-  // CK23: call void [[KERNEL:@.+]](i32* [[VAL]], float* [[_VAL]])
-  #pragma omp target is_device_ptr(tr,lr) defaultmap(none:pointer)
+// CK23-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK23-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK23-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK23-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK23-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK23-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK23-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK23-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
+// CK23-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
+// CK23-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
+// CK23-DAG: store i32* [[VAL]], i32** [[CP1]]
+// CK23-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
+// CK23-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
+
+// CK23-DAG: [[_BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
+// CK23-DAG: [[_P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
+// CK23-DAG: [[_CBP1:%.+]] = bitcast i8** [[_BP1]] to float**
+// CK23-DAG: [[_CP1:%.+]] = bitcast i8** [[_P1]] to float**
+// CK23-DAG: store float* [[_VAL:%.+]], float** [[_CBP1]]
+// CK23-DAG: store float* [[_VAL]], float** [[_CP1]]
+// CK23-DAG: [[_VAL]] = load float*, float** [[_ADDR:%.+]],
+// CK23-DAG: [[_ADDR]] = load float**, float*** [[_ADDR2:%.+]],
+
+// CK23: call void [[KERNEL:@.+]](i32* [[VAL]], float* [[_VAL]])
+#pragma omp target is_device_ptr(tr, lr) defaultmap(none \
+                                                    : pointer)
   {
     ++tr,++lr;
   }
@@ -1422,40 +1567,52 @@ void explicit_maps_single (int ii){
   // Map of a scalar.
   int a = ii;
 
-  // Close.
-  // Region 00
-  // CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}}, i8** null, i8** null)
-  // CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK24-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
-  // CK24-DAG: store i32* [[VAR0]], i32** [[CP0]]
-
-  // CK24: call void [[CALL00:@.+]](i32* {{[^,]+}})
-  #pragma omp target map(close, tofrom: a) defaultmap(none:scalar)
+// Close.
+// Region 00
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK24-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
+// CK24-DAG: store i32* [[VAR0]], i32** [[CP0]]
+
+// CK24: call void [[CALL00:@.+]](i32* {{[^,]+}})
+#pragma omp target map(close, tofrom        \
+                       : a) defaultmap(none \
+                                       : scalar)
   {
    a++;
   }
 
-  // Always Close.
-  // Region 01
-  // CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null, i8** null)
-  // CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK24-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
-  // CK24-DAG: store i32* [[VAR0]], i32** [[CP0]]
-
-  // CK24: call void [[CALL01:@.+]](i32* {{[^,]+}})
-  #pragma omp target map(always close tofrom: a) defaultmap(none:scalar)
+// Always Close.
+// Region 01
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK24-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK24-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK24-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
+// CK24-DAG: store i32* [[VAR0]], i32** [[CP0]]
+
+// CK24: call void [[CALL01:@.+]](i32* {{[^,]+}})
+#pragma omp target map(always close tofrom  \
+                       : a) defaultmap(none \
+                                       : scalar)
   {
    a++;
   }
@@ -1534,7 +1691,11 @@ void declare_target_link()
 #pragma omp target defaultmap(none:scalar) defaultmap(none:aggregate) defaultmap(none:pointer)
   {
 
-    // CK26-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 3, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
+    // CK26-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+    // CK26-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+    // CK26-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+    // CK26-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+    // CK26-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
     // CK26-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
     // CK26-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
     // CK26-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 2

diff --git a/clang/test/OpenMP/target_defaultmap_codegen_02.cpp b/clang/test/OpenMP/target_defaultmap_codegen_02.cpp
index 0c78661..c066347 100644 (file)
--- a/clang/test/OpenMP/target_defaultmap_codegen_02.cpp
+++ b/clang/test/OpenMP/target_defaultmap_codegen_02.cpp
@@ -27,16 +27,21 @@
 
 // CK1-LABEL: implicit_present_scalar{{.*}}(
 void implicit_present_scalar(int a) {
-  // CK1-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
-  // CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
-  // CK1-DAG: store i32* [[DECL:%[^,]+]], i32** [[CBP1]]
-  // CK1-DAG: store i32* [[DECL]], i32** [[CP1]]
-  #pragma omp target defaultmap(present: scalar)
+// CK1-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK1-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK1-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK1-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK1-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
+// CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
+// CK1-DAG: store i32* [[DECL:%[^,]+]], i32** [[CBP1]]
+// CK1-DAG: store i32* [[DECL]], i32** [[CP1]]
+#pragma omp target defaultmap(present \
+                              : scalar)
   {
     a += 1.0;
   }
@@ -71,18 +76,23 @@ void implicit_present_scalar(int a) {
 void implicit_present_aggregate(int a) {
   double darr[2] = {(double)a, (double)a};
 
-  // CK2-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK2-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK2-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK2-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK2-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK2-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [2 x double]**
-  // CK2-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [2 x double]**
-  // CK2-DAG: store [2 x double]* [[DECL:%[^,]+]], [2 x double]** [[CBP1]]
-  // CK2-DAG: store [2 x double]* [[DECL]], [2 x double]** [[CP1]]
-
-  // CK2: call void [[KERNEL:@.+]]([2 x double]* [[DECL]])
-  #pragma omp target defaultmap(present: aggregate)
+// CK2-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK2-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK2-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK2-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK2-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK2-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK2-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK2-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK2-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK2-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [2 x double]**
+// CK2-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [2 x double]**
+// CK2-DAG: store [2 x double]* [[DECL:%[^,]+]], [2 x double]** [[CBP1]]
+// CK2-DAG: store [2 x double]* [[DECL]], [2 x double]** [[CP1]]
+
+// CK2: call void [[KERNEL:@.+]]([2 x double]* [[DECL]])
+#pragma omp target defaultmap(present \
+                              : aggregate)
   {
     darr[0] += 1.0;
     darr[1] += 1.0;
@@ -119,9 +129,13 @@ void implicit_present_aggregate(int a) {
 void explicit_present_pointer() {
   int *pa;
 
-  // CK3-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE]]{{.+}}, i8** null, i8** null)
-  // CK3-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK3-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK3-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK3-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK3-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK3-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK3-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK3-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK3-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK3-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK3-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -165,18 +179,23 @@ void explicit_present_pointer() {
 void implicit_present_double_complex (int a){
   double _Complex dc = (double)a;
 
-  // CK4-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK4-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK4-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK4-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK4-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK4-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { double, double }**
-  // CK4-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { double, double }**
-  // CK4-DAG: store { double, double }* [[PTR:%[^,]+]], { double, double }** [[CBP1]]
-  // CK4-DAG: store { double, double }* [[PTR]], { double, double }** [[CP1]]
-
-  // CK4: call void [[KERNEL:@.+]]({ double, double }* [[PTR]])
-  #pragma omp target defaultmap(present:scalar)
+// CK4-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK4-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK4-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK4-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK4-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK4-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK4-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK4-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK4-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK4-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { double, double }**
+// CK4-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { double, double }**
+// CK4-DAG: store { double, double }* [[PTR:%[^,]+]], { double, double }** [[CBP1]]
+// CK4-DAG: store { double, double }* [[PTR]], { double, double }** [[CP1]]
+
+// CK4: call void [[KERNEL:@.+]]({ double, double }* [[PTR]])
+#pragma omp target defaultmap(present \
+                              : scalar)
   {
    dc *= dc;
   }

diff --git a/clang/test/OpenMP/target_depend_codegen.cpp b/clang/test/OpenMP/target_depend_codegen.cpp
index 1883d64..a9136b6 100644 (file)
--- a/clang/test/OpenMP/target_depend_codegen.cpp
+++ b/clang/test/OpenMP/target_depend_codegen.cpp
@@ -232,12 +232,12 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 0
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null)
-// CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
-// CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 -1, i32 0, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
+// CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
+// CHECK:       br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // CHECK:       [[FAIL]]
 // CHECK:       call void [[HVT0]]()
-// CHECK-NEXT:  br label %[[END]]
+// CHECK:       br label %[[END]]
 // CHECK:       [[END]]
 // CHECK:       ret i32 0
 
@@ -259,10 +259,10 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 2
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_nowait_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 3, i8** [[BPR]], i8** [[PR]], i64* [[SZT]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* [[MAPT]], i32 0, i32 0), i8** null, i8** [[M]], i32 0, i8* null, i32 0, i8* null)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 -1, i32 0, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
 
 // CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
-// CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
+// CHECK:       br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // CHECK:       [[FAIL]]
 // CHECK:       [[BP0:%.+]] = load i[[SZ]]*, i[[SZ]]** %
 // CHECK:       [[BP1_I32:%.+]] = load i32, i32* @
@@ -271,7 +271,7 @@ int foo(int n) {
 // CHECK-32:    store i32 [[BP1_I32]], i32* [[BP1_PTR:%.+]],
 // CHECK:       [[BP1:%.+]] = load i[[SZ]], i[[SZ]]* [[BP1_PTR]],
 // CHECK:       call void [[HVT1]](i[[SZ]]* [[BP0]], i[[SZ]] [[BP1]])
-// CHECK-NEXT:  br label %[[END]]
+// CHECK:       br label %[[END]]
 // CHECK:       [[END]]
 // CHECK:       ret i32 0
 

diff --git a/clang/test/OpenMP/target_device_codegen.cpp b/clang/test/OpenMP/target_device_codegen.cpp
index 4839f86..58c3976 100644 (file)
--- a/clang/test/OpenMP/target_device_codegen.cpp
+++ b/clang/test/OpenMP/target_device_codegen.cpp
@@ -27,38 +27,40 @@
 
 void foo(int n) {
 
-  // CHECK:       [[N:%.+]] = load i32, i32* [[N_ADDR:%.+]],
-  // CHECK:       store i32 [[N]], i32* [[DEVICE_CAP:%.+]],
-  // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
-  // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-  // CHECK:       [[RET:%.+]] = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null)
-  // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
-  // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
-  // CHECK:       [[FAIL]]
-  // CHECK:       call void [[HVT0:@.+]]()
-  // CHECK-NEXT:  br label %[[END]]
-  // CHECK:       [[END]]
-  #pragma omp target device(n)
+// CHECK:       [[N:%.+]] = load i32, i32* [[N_ADDR:%.+]],
+// CHECK:       store i32 [[N]], i32* [[DEVICE_CAP:%.+]],
+// CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
+// CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 -1, i32 0, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
+// CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
+// CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
+// CHECK:       [[FAIL]]
+// CHECK:       call void [[HVT0:@.+]]()
+// CHECK-NEXT:  br label %[[END]]
+// CHECK:       [[END]]
+#pragma omp target device(n)
   ;
-  // CHECK:       [[N:%.+]] = load i32, i32* [[N_ADDR]],
-  // CHECK:       store i32 [[N]], i32* [[DEVICE_CAP:%.+]],
-  // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
-  // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-  // CHECK:       [[RET:%.+]] = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null)
-  // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
-  // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
-  // CHECK:       [[FAIL]]
-  // CHECK:       call void [[HVT0:@.+]]()
-  // CHECK-NEXT:  br label %[[END]]
-  // CHECK:       [[END]]
-  #pragma omp target device(device_num: n)
+// CHECK:       [[N:%.+]] = load i32, i32* [[N_ADDR]],
+// CHECK:       store i32 [[N]], i32* [[DEVICE_CAP:%.+]],
+// CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
+// CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 -1, i32 0, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
+// CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
+// CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
+// CHECK:       [[FAIL]]
+// CHECK:       call void [[HVT0:@.+]]()
+// CHECK-NEXT:  br label %[[END]]
+// CHECK:       [[END]]
+#pragma omp target device(device_num \
+                          : n)
   ;
 
 #ifdef OMP99
-  // REV-NOT:   call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}},
-  // REV:       call void @__omp_offloading_{{.+}}_l61()
-  // REV-NOT:   call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}},
-  #pragma omp target device(ancestor: n)
+// REV-NOT:   call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}},
+// REV:       call void @__omp_offloading_{{.+}}_l62()
+// REV-NOT:   call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}},
+#pragma omp target device(ancestor \
+                          : n)
   ;
 #endif
 }

diff --git a/clang/test/OpenMP/target_firstprivate_codegen.cpp b/clang/test/OpenMP/target_firstprivate_codegen.cpp
index d4252e2..8fb6c7b 100644 (file)
--- a/clang/test/OpenMP/target_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/target_firstprivate_codegen.cpp
@@ -141,7 +141,7 @@ int foo(int n, double *ptr) {
   // CHECK:  store i{{64|32}} [[GA_VAL]], i{{64|32}}* [[PCAST_TOPTR2]],
   // CHECK:  [[BASE_PTR_GEP_ARG:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[BASE_PTR_ARR]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
   // CHECK:  [[PTR_GEP_ARG:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[PTR_ARR]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
-  // CHECK:  {{.+}} = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, {{.+}}, i32 3, i8** [[BASE_PTR_GEP_ARG]], i8** [[PTR_GEP_ARG]], i[[SZ]]* getelementptr inbounds ([3 x i[[SZ]]], [3 x i[[SZ]]]* [[SIZET]], i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* [[MAPT]], i32 0, i32 0), i8** null, i8** null)
+  // CHECK:  {{.+}} = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
 
   // TCHECK:  define weak_odr void @__omp_offloading_{{.+}}(i{{[0-9]+}} noundef [[A_IN:%.+]], i32** noundef nonnull align {{[0-9]+}} dereferenceable({{[0-9]+}}) [[P_IN:%.+]], i{{[0-9]+}} noundef [[GA_IN:%.+]])
   // TCHECK:  [[A_ADDR:%.+]] = alloca i{{[0-9]+}},
@@ -253,7 +253,7 @@ int foo(int n, double *ptr) {
   // CHECK:  [[BASE_PTR_GEP_ARG2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BASE_PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
   // CHECK:  [[PTR_GEP_ARG2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
   // CHECK:  [[SIZES_ARG2:%.+]] = getelementptr inbounds [9 x i[[SZ]]], [9 x i[[SZ]]]* [[SIZET2]],  i{{[0-9]+}} 0, i{{[0-9]+}} 0
-  // CHECK: {{.+}} = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, {{.+}}, i32 9, i8** [[BASE_PTR_GEP_ARG2]], i8** [[PTR_GEP_ARG2]], i[[SZ]]* [[SIZES_ARG2]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* [[MAPT2]], i32 0, i32 0), i8** null, i8** null)
+  // CHECK:  {{.+}} = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
 
   // make sure that firstprivate variables are generated in all cases and that we use those instances for operations inside the
   // target region
@@ -350,7 +350,7 @@ int foo(int n, double *ptr) {
 
   // CHECK:  [[BASE_PTR_GEP_ARG3:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[BASE_PTR_ARR3]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
   // CHECK:  [[PTR_GEP_ARG3:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[PTR_ARR3]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
-  // CHECK: {{.+}} = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, {{.+}}, i32 2, i8** [[BASE_PTR_GEP_ARG3]], i8** [[PTR_GEP_ARG3]], i[[SZ]]* getelementptr inbounds ([2 x i[[SZ]]], [2 x i[[SZ]]]* [[SIZET3]], i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT3]], i32 0, i32 0), i8** null, i8** null)
+  // CHECK:  {{.+}} = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
 
   // TCHECK:  define weak_odr void @__omp_offloading_{{.+}}(double* noundef [[PTR_IN:%.+]], [[TTII]]* noundef nonnull align {{[0-9]+}} dereferenceable({{[0-9]+}}) [[E:%.+]])
   // TCHECK:  [[PTR_ADDR:%.+]] = alloca double*,
@@ -477,7 +477,7 @@ struct S1 {
   // CHECK:  store i{{[0-9]+}} [[B_SIZE:%.+]], i{{[0-9]+}}* [[SIZES_GEP4_4]],
 
   // only check that we use the map types stored in the global variable
-  // CHECK:  call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, {{.+}}, i32 5, i8** {{.+}}, i8** {{.+}}, i{{[0-9]+}}* {{.+}}, i64* getelementptr inbounds ([5 x i64], [5 x i64]* [[MAPT4]], i32 0, i32 0), i8** null, i8** null)
+  // CHECK:  {{.+}} = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
 
   // TCHECK: define weak_odr void @__omp_offloading_{{.+}}([[S1]]* noundef [[TH:%.+]], i{{[0-9]+}} noundef [[B_IN:%.+]], i{{[0-9]+}} noundef [[VLA:%.+]], i{{[0-9]+}} noundef [[VLA1:%.+]], i{{[0-9]+}}{{.+}} [[C_IN:%.+]])
   // TCHECK:  [[TH_ADDR:%.+]] = alloca [[S1]]*,
@@ -549,7 +549,7 @@ struct S1 {
   // CHECK:  store [10 x i{{[0-9]+}}]* [[B]], [10 x i{{[0-9]+}}]** [[BCAST_TOPTR]],
 
   // only check that the right sizes and map types are used
-  // CHECK:  call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, {{.+}}, i32 3, i8** {{.+}}, i8** {{.+}}, i[[SZ]]* getelementptr inbounds ([3 x i[[SZ]]], [3 x i[[SZ]]]* [[SIZET5]], i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* [[MAPT5]], i32 0, i32 0), i8** null, i8** null)
+  // CHECK:  {{.+}} = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
 };
 
 int bar(int n, double *ptr) {
@@ -585,7 +585,7 @@ int bar(int n, double *ptr) {
 // CHECK:  [[BCAST_TOPTR:%.+]] = bitcast i8** [[PTRS_GEP6_1]] to [10 x i{{[0-9]+}}]**
 // CHECK:  store [10 x i{{[0-9]+}}]* [[B]], [10 x i{{[0-9]+}}]** [[BCAST_TOPTR]],
 
-// CHECK:  call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, {{.+}}, i32 2, i8** {{.+}}, i8** {{.+}}, i[[SZ]]* getelementptr inbounds ([2 x i[[SZ]]], [2 x i[[SZ]]]* [[SIZET6]], i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT6]], i32 0, i32 0), i8** null, i8** null)
+// CHECK:  {{.+}} = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
 
 // TCHECK: define weak_odr void @__omp_offloading_{{.+}}(i{{[0-9]+}} noundef [[A_IN:%.+]], [10 x i{{[0-9]+}}]*{{.+}} [[B_IN:%.+]])
 // TCHECK:  [[A_ADDR:%.+]] = alloca i{{[0-9]+}},

diff --git a/clang/test/OpenMP/target_is_device_ptr_codegen.cpp b/clang/test/OpenMP/target_is_device_ptr_codegen.cpp
index 8ba2438..ee09c1b 100644 (file)
--- a/clang/test/OpenMP/target_is_device_ptr_codegen.cpp
+++ b/clang/test/OpenMP/target_is_device_ptr_codegen.cpp
@@ -49,134 +49,162 @@ void foo(float *&lr, T *&tr) {
   float *l;
   T *t;
 
-  // CK1-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES00]]{{.+}}, {{.+}}[[TYPES00]]{{.+}}, i8** null, i8** null)
-  // CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
-  // CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
-  // CK1-DAG: store double* [[VAL:%.+]], double** [[CBP1]]
-  // CK1-DAG: store double* [[VAL]], double** [[CP1]]
-  // CK1-DAG: [[VAL]] = load double*, double** [[ADDR:@g]],
-
-  // CK1: call void [[KERNEL:@.+]](double* [[VAL]])
-  #pragma omp target is_device_ptr(g)
+// CK1-DAG: [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK1-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK1-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK1-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK1-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
+// CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
+// CK1-DAG: store double* [[VAL:%.+]], double** [[CBP1]]
+// CK1-DAG: store double* [[VAL]], double** [[CP1]]
+// CK1-DAG: [[VAL]] = load double*, double** [[ADDR:@g]],
+
+// CK1: call void [[KERNEL:@.+]](double* [[VAL]])
+#pragma omp target is_device_ptr(g)
   {
     ++g;
   }
 
-  // CK1-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES01]]{{.+}}, {{.+}}[[TYPES01]]{{.+}}, i8** null, i8** null)
-  // CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
-  // CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
-  // CK1-DAG: store float* [[VAL:%.+]], float** [[CBP1]]
-  // CK1-DAG: store float* [[VAL]], float** [[CP1]]
-  // CK1-DAG: [[VAL]] = load float*, float** [[ADDR:%.+]],
-
-  // CK1: call void [[KERNEL:@.+]](float* [[VAL]])
-  #pragma omp target is_device_ptr(l)
+// CK1-DAG: [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK1-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK1-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK1-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK1-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
+// CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
+// CK1-DAG: store float* [[VAL:%.+]], float** [[CBP1]]
+// CK1-DAG: store float* [[VAL]], float** [[CP1]]
+// CK1-DAG: [[VAL]] = load float*, float** [[ADDR:%.+]],
+
+// CK1: call void [[KERNEL:@.+]](float* [[VAL]])
+#pragma omp target is_device_ptr(l)
   {
     ++l;
   }
 
-  // CK1-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES02]]{{.+}}, {{.+}}[[TYPES02]]{{.+}}, i8** null, i8** null)
-  // CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
-  // CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
-  // CK1-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
-  // CK1-DAG: store i32* [[VAL]], i32** [[CP1]]
-  // CK1-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
-
-  // CK1: call void [[KERNEL:@.+]](i32* [[VAL]])
-  #pragma omp target is_device_ptr(t)
+// CK1-DAG: [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK1-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK1-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK1-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK1-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
+// CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
+// CK1-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
+// CK1-DAG: store i32* [[VAL]], i32** [[CP1]]
+// CK1-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
+
+// CK1: call void [[KERNEL:@.+]](i32* [[VAL]])
+#pragma omp target is_device_ptr(t)
   {
     ++t;
   }
 
-  // CK1-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES03]]{{.+}}, {{.+}}[[TYPES03]]{{.+}}, i8** null, i8** null)
-  // CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
-  // CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
-  // CK1-DAG: store float* [[VAL:%.+]], float** [[CBP1]]
-  // CK1-DAG: store float* [[VAL]], float** [[CP1]]
-  // CK1-DAG: [[VAL]] = load float*, float** [[ADDR:%.+]],
-  // CK1-DAG: [[ADDR]] = load float**, float*** [[ADDR2:%.+]],
-
-  // CK1: call void [[KERNEL:@.+]](float* [[VAL]])
-  #pragma omp target is_device_ptr(lr)
+// CK1-DAG: [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK1-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK1-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK1-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK1-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
+// CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
+// CK1-DAG: store float* [[VAL:%.+]], float** [[CBP1]]
+// CK1-DAG: store float* [[VAL]], float** [[CP1]]
+// CK1-DAG: [[VAL]] = load float*, float** [[ADDR:%.+]],
+// CK1-DAG: [[ADDR]] = load float**, float*** [[ADDR2:%.+]],
+
+// CK1: call void [[KERNEL:@.+]](float* [[VAL]])
+#pragma omp target is_device_ptr(lr)
   {
     ++lr;
   }
 
-  // CK1-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES04]]{{.+}}, {{.+}}[[TYPES04]]{{.+}}, i8** null, i8** null)
-  // CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
-  // CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
-  // CK1-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
-  // CK1-DAG: store i32* [[VAL]], i32** [[CP1]]
-  // CK1-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
-  // CK1-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
-
-  // CK1: call void [[KERNEL:@.+]](i32* [[VAL]])
-  #pragma omp target is_device_ptr(tr)
+// CK1-DAG: [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK1-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK1-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK1-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK1-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
+// CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
+// CK1-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
+// CK1-DAG: store i32* [[VAL]], i32** [[CP1]]
+// CK1-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
+// CK1-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
+
+// CK1: call void [[KERNEL:@.+]](i32* [[VAL]])
+#pragma omp target is_device_ptr(tr)
   {
     ++tr;
   }
 
-  // CK1-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES05]]{{.+}}, {{.+}}[[TYPES05]]{{.+}}, i8** null, i8** null)
-  // CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
-  // CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
-  // CK1-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
-  // CK1-DAG: store i32* [[VAL]], i32** [[CP1]]
-  // CK1-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
-  // CK1-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
-
-  // CK1: call void [[KERNEL:@.+]](i32* [[VAL]])
-  #pragma omp target is_device_ptr(tr,lr)
+// CK1-DAG: [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK1-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK1-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK1-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK1-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
+// CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
+// CK1-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
+// CK1-DAG: store i32* [[VAL]], i32** [[CP1]]
+// CK1-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
+// CK1-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
+
+// CK1: call void [[KERNEL:@.+]](i32* [[VAL]])
+#pragma omp target is_device_ptr(tr, lr)
   {
     ++tr;
   }
 
-  // CK1-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 2, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES06]]{{.+}}, {{.+}}[[TYPES06]]{{.+}}, i8** null, i8** null)
-  // CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
-  // CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
-  // CK1-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
-  // CK1-DAG: store i32* [[VAL]], i32** [[CP1]]
-  // CK1-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
-  // CK1-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
-
-  // CK1-DAG: [[_BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
-  // CK1-DAG: [[_P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
-  // CK1-DAG: [[_CBP1:%.+]] = bitcast i8** [[_BP1]] to float**
-  // CK1-DAG: [[_CP1:%.+]] = bitcast i8** [[_P1]] to float**
-  // CK1-DAG: store float* [[_VAL:%.+]], float** [[_CBP1]]
-  // CK1-DAG: store float* [[_VAL]], float** [[_CP1]]
-  // CK1-DAG: [[_VAL]] = load float*, float** [[_ADDR:%.+]],
-  // CK1-DAG: [[_ADDR]] = load float**, float*** [[_ADDR2:%.+]],
-
-  // CK1: call void [[KERNEL:@.+]](i32* [[VAL]], float* [[_VAL]])
-  #pragma omp target is_device_ptr(tr,lr)
+// CK1-DAG: [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK1-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK1-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK1-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK1-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
+// CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
+// CK1-DAG: store i32* [[VAL:%.+]], i32** [[CBP1]]
+// CK1-DAG: store i32* [[VAL]], i32** [[CP1]]
+// CK1-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
+// CK1-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
+
+// CK1-DAG: [[_BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
+// CK1-DAG: [[_P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
+// CK1-DAG: [[_CBP1:%.+]] = bitcast i8** [[_BP1]] to float**
+// CK1-DAG: [[_CP1:%.+]] = bitcast i8** [[_P1]] to float**
+// CK1-DAG: store float* [[_VAL:%.+]], float** [[_CBP1]]
+// CK1-DAG: store float* [[_VAL]], float** [[_CP1]]
+// CK1-DAG: [[_VAL]] = load float*, float** [[_ADDR:%.+]],
+// CK1-DAG: [[_ADDR]] = load float**, float*** [[_ADDR2:%.+]],
+
+// CK1: call void [[KERNEL:@.+]](i32* [[VAL]], float* [[_VAL]])
+#pragma omp target is_device_ptr(tr, lr)
   {
     ++tr,++lr;
   }
@@ -206,21 +234,6 @@ void bar(float *&a, int *&b) {
 
 // CK2: [[ST:%.+]] = type { double*, double** }
 
-// CK2-LABEL: @.__omp_offloading_{{.*}}foo{{.*}}_l244.region_id = weak{{.*}} constant i8 0
-
-// CK2: [[SIZE00:@.+]] = {{.+}}constant [1 x i[[sz:64|32]]] [i{{64|32}} {{8|4}}]
-// CK2: [[MTYPE00:@.+]] = {{.+}}constant [1 x i64] [i64 288]
-
-// CK2-LABEL: @.__omp_offloading_{{.*}}foo{{.*}}_l259.region_id = weak{{.*}} constant i8 0
-
-// CK2: [[SIZE01:@.+]] = {{.+}}constant [1 x i[[sz]]] [i[[sz]] {{8|4}}]
-// CK2: [[MTYPE01:@.+]] = {{.+}}constant [1 x i64] [i64 288]
-
-// CK2-LABEL: @.__omp_offloading_{{.*}}foo{{.*}}_l274.region_id = weak{{.*}} constant i8 0
-
-// CK2: [[SIZE02:@.+]] = {{.+}}constant [1 x i[[sz]]] [i[[sz]] {{8|4}}]
-// CK2: [[MTYPE02:@.+]] = {{.+}}constant [1 x i64] [i64 288]
-
 template <typename T>
 struct ST {
   T *a;
@@ -231,47 +244,59 @@ struct ST {
   void foo(double *&arg) {
     int *la = 0;
 
-    // CK2-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}}, i8** null, i8** null)
-    // CK2-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK2-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK2-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK2-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK2-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
-    // CK2-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
-    // CK2-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
-    // CK2-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
-    #pragma omp target is_device_ptr(a)
+// CK2-DAG: [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK2-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK2-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK2-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK2-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK2-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK2-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK2-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK2-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK2-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
+// CK2-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
+// CK2-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
+// CK2-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
+#pragma omp target is_device_ptr(a)
     {
       a++;
     }
 
-    // CK2-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null, i8** null)
-    // CK2-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK2-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK2-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK2-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK2-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
-    // CK2-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
-    // CK2-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
-    // CK2-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
-    #pragma omp target is_device_ptr(b)
+// CK2-DAG: [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK2-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK2-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK2-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK2-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK2-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK2-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK2-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK2-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK2-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
+// CK2-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
+// CK2-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
+// CK2-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
+#pragma omp target is_device_ptr(b)
     {
       b++;
     }
 
-    // CK2-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE02]]{{.+}}, i8** null, i8** null)
-    // CK2-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK2-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK2-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK2-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK2-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
-    // CK2-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
-    // CK2-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
-    // CK2-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
-    #pragma omp target is_device_ptr(a, b)
+// CK2-DAG: [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK2-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK2-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK2-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK2-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK2-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK2-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK2-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK2-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK2-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
+// CK2-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
+// CK2-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
+// CK2-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
+#pragma omp target is_device_ptr(a, b)
     {
       a++;
       b++;
@@ -307,7 +332,11 @@ void bar(double *arg){
 // CK3-DAG: [[TYPES:@.+]] = {{.+}}constant [1 x i64] [i64 [[#0x21]]]
 void bar() {
   __attribute__((aligned(64))) double *ptr;
-  // CK3-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
+  // CK3-DAG: [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK3-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK3-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK3-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK3-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
   // CK3-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
   // CK3-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
   // CK3-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0

diff --git a/clang/test/OpenMP/target_map_codegen_00.cpp b/clang/test/OpenMP/target_map_codegen_00.cpp
index 6925c76..bfc55af 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_00.cpp
+++ b/clang/test/OpenMP/target_map_codegen_00.cpp
@@ -73,21 +73,25 @@ void implicit_maps_integer (int a){
   B::modify(a);
   int i = a;
 
-  // CK1-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
-  // CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
-  // CK1-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
-  // CK1-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
-  // CK1-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
-  // CK1-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
-  // CK1-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
-
-  // CK1: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
-  #pragma omp target
+// CK1-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK1-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK1-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK1-DAG: [[PGEP:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK1-DAG: store i8** [[PGEP:%.+]], i8*** [[BPARG]]
+// CK1-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK1-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK1-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK1-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
+// CK1-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
+// CK1-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
+// CK1-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
+// CK1-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
+// CK1-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
+// CK1-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
+
+// CK1: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
+#pragma omp target
   {
    ++i;
   }

diff --git a/clang/test/OpenMP/target_map_codegen_01.cpp b/clang/test/OpenMP/target_map_codegen_01.cpp
index c275ca4..42ead5e 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_01.cpp
+++ b/clang/test/OpenMP/target_map_codegen_01.cpp
@@ -48,40 +48,48 @@
 // CK2-LABEL: implicit_maps_reference{{.*}}(
 void implicit_maps_reference (int a, int *b){
   int &i = a;
-  // CK2-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK2-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK2-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK2-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK2-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK2-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
-  // CK2-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
-  // CK2-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
-  // CK2-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
-  // CK2-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
-  // CK2-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
-  // CK2-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
+// CK2-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK2-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK2-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK2-DAG: [[PGEP:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK2-DAG: store i8** [[PGEP:%.+]], i8*** [[BPARG]]
+// CK2-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK2-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK2-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK2-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK2-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
+// CK2-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
+// CK2-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
+// CK2-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
+// CK2-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
+// CK2-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
+// CK2-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
 
-  // CK2: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
-  #pragma omp target
+// CK2: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
+#pragma omp target
   {
    ++i;
   }
 
   int *&p = b;
-  // CK2-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES2]]{{.+}}, {{.+}}[[TYPES2]]{{.+}}, i8** null, i8** null)
-  // CK2-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK2-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK2-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK2-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK2-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
-  // CK2-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
-  // CK2-DAG: store i32* [[VAL:%[^,]+]], i32** [[CBP1]]
-  // CK2-DAG: store i32* [[VAL]], i32** [[CP1]]
-  // CK2-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
-  // CK2-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
+// CK2-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK2-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK2-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK2-DAG: [[PGEP:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK2-DAG: store i8** [[PGEP:%.+]], i8*** [[BPARG]]
+// CK2-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK2-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK2-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK2-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK2-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
+// CK2-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
+// CK2-DAG: store i32* [[VAL:%[^,]+]], i32** [[CBP1]]
+// CK2-DAG: store i32* [[VAL]], i32** [[CP1]]
+// CK2-DAG: [[VAL]] = load i32*, i32** [[ADDR:%.+]],
+// CK2-DAG: [[ADDR]] = load i32**, i32*** [[ADDR2:%.+]],
 
-  // CK2: call void [[KERNEL2:@.+]](i32* [[VAL]])
-  #pragma omp target
+// CK2: call void [[KERNEL2:@.+]](i32* [[VAL]])
+#pragma omp target
   {
    ++p;
   }

diff --git a/clang/test/OpenMP/target_map_codegen_02.cpp b/clang/test/OpenMP/target_map_codegen_02.cpp
index d012e3e..e885fc0 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_02.cpp
+++ b/clang/test/OpenMP/target_map_codegen_02.cpp
@@ -42,21 +42,25 @@
 // CK3-LABEL: implicit_maps_parameter{{.*}}(
 void implicit_maps_parameter (int a){
 
-  // CK3-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK3-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK3-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK3-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK3-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK3-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
-  // CK3-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
-  // CK3-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
-  // CK3-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
-  // CK3-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
-  // CK3-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
-  // CK3-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
+// CK3-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK3-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK3-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK3-DAG: [[PGEP:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK3-DAG: store i8** [[PGEP:%.+]], i8*** [[BPARG]]
+// CK3-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK3-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK3-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK3-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK3-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
+// CK3-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
+// CK3-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
+// CK3-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
+// CK3-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
+// CK3-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
+// CK3-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
 
-  // CK3: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
-  #pragma omp target
+// CK3: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
+#pragma omp target
   {
    ++a;
   }

diff --git a/clang/test/OpenMP/target_map_codegen_03.cpp b/clang/test/OpenMP/target_map_codegen_03.cpp
index 0606033..36d9de4 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_03.cpp
+++ b/clang/test/OpenMP/target_map_codegen_03.cpp
@@ -97,9 +97,26 @@ void implicit_maps_nested_integer (int a){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP7]], align 8
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[TMP10]], 0
-// CHECK1-NEXT:    br i1 [[TMP11]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP9]], i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i64 [[TMP2]]) #[[ATTR3:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -179,9 +196,26 @@ void implicit_maps_nested_integer (int a){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP7]], align 4
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[TMP10]], 0
-// CHECK3-NEXT:    br i1 [[TMP11]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP9]], i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK3-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i32 [[TMP2]]) #[[ATTR3:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/target_map_codegen_04.cpp b/clang/test/OpenMP/target_map_codegen_04.cpp
index 79cb9b8..f2c0c8f 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_04.cpp
+++ b/clang/test/OpenMP/target_map_codegen_04.cpp
@@ -48,21 +48,25 @@ void implicit_maps_nested_integer_and_enum (int a){
   // Using an enum should not change the mapping information.
   int  i = a;
 
-  // CK5-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK5-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK5-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK5-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK5-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK5-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
-  // CK5-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
-  // CK5-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
-  // CK5-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
-  // CK5-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
-  // CK5-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
-  // CK5-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
+// CK5-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK5-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK5-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK5-DAG: [[PGEP:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK5-DAG: store i8** [[PGEP:%.+]], i8*** [[BPARG]]
+// CK5-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK5-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK5-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK5-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK5-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
+// CK5-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
+// CK5-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
+// CK5-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
+// CK5-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
+// CK5-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
+// CK5-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
 
-  // CK5: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
-  #pragma omp target
+// CK5: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
+#pragma omp target
   {
     ++i;
     i += SomeEnum;

diff --git a/clang/test/OpenMP/target_map_codegen_05.cpp b/clang/test/OpenMP/target_map_codegen_05.cpp
index 8fd609c..6de1c31 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_05.cpp
+++ b/clang/test/OpenMP/target_map_codegen_05.cpp
@@ -42,23 +42,27 @@
 // CK6-LABEL: implicit_maps_host_global{{.*}}(
 int Gi;
 void implicit_maps_host_global (int a){
-  // CK6-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK6-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK6-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK6-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK6-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK6-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
-  // CK6-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
-  // CK6-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
-  // CK6-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
-  // CK6-64-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
-  // CK6-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
-  // CK6-64-DAG: store i32 [[GBLVAL:%.+]], i32* [[CADDR]],
-  // CK6-64-DAG: [[GBLVAL]] = load i32, i32* [[GBL]],
-  // CK6-32-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[GBLVAL:%.+]],
+// CK6-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK6-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK6-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK6-DAG: [[PGEP:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK6-DAG: store i8** [[PGEP:%.+]], i8*** [[BPARG]]
+// CK6-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK6-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK6-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK6-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK6-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
+// CK6-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
+// CK6-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
+// CK6-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
+// CK6-64-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
+// CK6-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to i32*
+// CK6-64-DAG: store i32 [[GBLVAL:%.+]], i32* [[CADDR]],
+// CK6-64-DAG: [[GBLVAL]] = load i32, i32* [[GBL]],
+// CK6-32-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[GBLVAL:%.+]],
 
-  // CK6: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
-  #pragma omp target
+// CK6: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
+#pragma omp target
   {
     ++Gi;
   }

diff --git a/clang/test/OpenMP/target_map_codegen_06.cpp b/clang/test/OpenMP/target_map_codegen_06.cpp
index 4d710fe..7c30c61 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_06.cpp
+++ b/clang/test/OpenMP/target_map_codegen_06.cpp
@@ -48,28 +48,32 @@
 void implicit_maps_double (int a){
   double d = (double)a;
 
-  // CK7-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK7-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK7-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK7-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK7-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK7-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK7-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK7-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK7-DAG: [[PGEP:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK7-DAG: store i8** [[PGEP:%.+]], i8*** [[BPARG]]
+// CK7-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK7-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK7-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK7-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
 
-  // CK7-64-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
-  // CK7-64-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
-  // CK7-64-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
-  // CK7-64-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
-  // CK7-64-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
-  // CK7-64-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to double*
-  // CK7-64-64-DAG: store double {{.+}}, double* [[CADDR]],
+// CK7-64-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
+// CK7-64-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
+// CK7-64-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
+// CK7-64-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
+// CK7-64-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
+// CK7-64-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to double*
+// CK7-64-64-DAG: store double {{.+}}, double* [[CADDR]],
 
-  // CK7-32-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
-  // CK7-32-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
-  // CK7-32-DAG: store double* [[DECL:%[^,]+]], double** [[CBP1]]
-  // CK7-32-DAG: store double* [[DECL]], double** [[CP1]]
+// CK7-32-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
+// CK7-32-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
+// CK7-32-DAG: store double* [[DECL:%[^,]+]], double** [[CBP1]]
+// CK7-32-DAG: store double* [[DECL]], double** [[CP1]]
 
-  // CK7-64: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
-  // CK7-32: call void [[KERNEL:@.+]](double* [[DECL]])
-  #pragma omp target
+// CK7-64: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
+// CK7-32: call void [[KERNEL:@.+]](double* [[DECL]])
+#pragma omp target
   {
     d += 1.0;
   }

diff --git a/clang/test/OpenMP/target_map_codegen_07.cpp b/clang/test/OpenMP/target_map_codegen_07.cpp
index f8e63d6..7ed7378 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_07.cpp
+++ b/clang/test/OpenMP/target_map_codegen_07.cpp
@@ -43,21 +43,25 @@
 void implicit_maps_float (int a){
   float f = (float)a;
 
-  // CK8-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK8-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK8-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK8-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK8-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK8-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
-  // CK8-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
-  // CK8-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
-  // CK8-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
-  // CK8-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
-  // CK8-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to float*
-  // CK8-DAG: store float {{.+}}, float* [[CADDR]],
+// CK8-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK8-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK8-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK8-DAG: [[PGEP:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK8-DAG: store i8** [[PGEP:%.+]], i8*** [[BPARG]]
+// CK8-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK8-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK8-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK8-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK8-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
+// CK8-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
+// CK8-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
+// CK8-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
+// CK8-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
+// CK8-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to float*
+// CK8-DAG: store float {{.+}}, float* [[CADDR]],
 
-  // CK8: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
-  #pragma omp target
+// CK8: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
+#pragma omp target
   {
     f += 1.0;
   }

diff --git a/clang/test/OpenMP/target_map_codegen_08.cpp b/clang/test/OpenMP/target_map_codegen_08.cpp
index b0b1c92..66669e6 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_08.cpp
+++ b/clang/test/OpenMP/target_map_codegen_08.cpp
@@ -43,18 +43,22 @@
 void implicit_maps_array (int a){
   double darr[2] = {(double)a, (double)a};
 
-  // CK9-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK9-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK9-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK9-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK9-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK9-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [2 x double]**
-  // CK9-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [2 x double]**
-  // CK9-DAG: store [2 x double]* [[DECL:%[^,]+]], [2 x double]** [[CBP1]]
-  // CK9-DAG: store [2 x double]* [[DECL]], [2 x double]** [[CP1]]
+// CK9-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK9-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK9-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK9-DAG: [[PGEP:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK9-DAG: store i8** [[PGEP:%.+]], i8*** [[BPARG]]
+// CK9-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK9-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK9-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK9-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK9-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [2 x double]**
+// CK9-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [2 x double]**
+// CK9-DAG: store [2 x double]* [[DECL:%[^,]+]], [2 x double]** [[CBP1]]
+// CK9-DAG: store [2 x double]* [[DECL]], [2 x double]** [[CP1]]
 
-  // CK9: call void [[KERNEL:@.+]]([2 x double]* [[DECL]])
-  #pragma omp target
+// CK9: call void [[KERNEL:@.+]]([2 x double]* [[DECL]])
+#pragma omp target
   {
     darr[0] += 1.0;
     darr[1] += 1.0;

diff --git a/clang/test/OpenMP/target_map_codegen_09.cpp b/clang/test/OpenMP/target_map_codegen_09.cpp
index 090dff4..a83869f 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_09.cpp
+++ b/clang/test/OpenMP/target_map_codegen_09.cpp
@@ -43,18 +43,22 @@
 void implicit_maps_pointer (){
   double *ddyn;
 
-  // CK10-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK10-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK10-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK10-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK10-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK10-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
-  // CK10-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
-  // CK10-DAG: store double* [[PTR:%[^,]+]], double** [[CBP1]]
-  // CK10-DAG: store double* [[PTR]], double** [[CP1]]
+// CK10-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK10-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK10-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK10-DAG: [[PGEP:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK10-DAG: store i8** [[PGEP:%.+]], i8*** [[BPARG]]
+// CK10-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK10-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK10-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK10-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK10-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to double**
+// CK10-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double**
+// CK10-DAG: store double* [[PTR:%[^,]+]], double** [[CBP1]]
+// CK10-DAG: store double* [[PTR]], double** [[CP1]]
 
-  // CK10: call void [[KERNEL:@.+]](double* [[PTR]])
-  #pragma omp target
+// CK10: call void [[KERNEL:@.+]](double* [[PTR]])
+#pragma omp target
   {
     ddyn[0] += 1.0;
     ddyn[1] += 1.0;

diff --git a/clang/test/OpenMP/target_map_codegen_10.cpp b/clang/test/OpenMP/target_map_codegen_10.cpp
index b02c089..eb10689 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_10.cpp
+++ b/clang/test/OpenMP/target_map_codegen_10.cpp
@@ -29,18 +29,23 @@
 void implicit_maps_double_complex (int a, int *b){
   double _Complex dc = (double)a;
 
-  // CK11-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 2, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK11-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK11-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK11-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK11-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK11-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { double, double }**
-  // CK11-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { double, double }**
-  // CK11-DAG: store { double, double }* [[PTR:%[^,]+]], { double, double }** [[CBP1]]
-  // CK11-DAG: store { double, double }* [[PTR]], { double, double }** [[CP1]]
-
-  // CK11: call void [[KERNEL:@.+]]({ double, double }* [[PTR]], i32** %{{.+}})
-  #pragma omp target defaultmap(tofrom:scalar)
+// CK11-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK11-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK11-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK11-DAG: [[PGEP:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK11-DAG: store i8** [[PGEP:%.+]], i8*** [[BPARG]]
+// CK11-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK11-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK11-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK11-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK11-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { double, double }**
+// CK11-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { double, double }**
+// CK11-DAG: store { double, double }* [[PTR:%[^,]+]], { double, double }** [[CBP1]]
+// CK11-DAG: store { double, double }* [[PTR]], { double, double }** [[CP1]]
+
+// CK11: call void [[KERNEL:@.+]]({ double, double }* [[PTR]], i32** %{{.+}})
+#pragma omp target defaultmap(tofrom \
+                              : scalar)
   {
    dc *= dc; *b = 1;
   }

diff --git a/clang/test/OpenMP/target_map_codegen_11.cpp b/clang/test/OpenMP/target_map_codegen_11.cpp
index 5712c22..133cc83 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_11.cpp
+++ b/clang/test/OpenMP/target_map_codegen_11.cpp
@@ -48,28 +48,32 @@
 void implicit_maps_float_complex (int a){
   float _Complex fc = (float)a;
 
-  // CK12-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK12-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK12-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK12-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK12-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK12-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK12-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK12-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK12-DAG: [[PGEP:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK12-DAG: store i8** [[PGEP:%.+]], i8*** [[BPARG]]
+// CK12-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK12-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK12-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK12-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
 
-  // CK12-64-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
-  // CK12-64-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
-  // CK12-64-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
-  // CK12-64-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
-  // CK12-64-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
-  // CK12-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to { float, float }*
-  // CK12-64-DAG: store { float, float } {{.+}}, { float, float }* [[CADDR]],
+// CK12-64-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz:64|32]]*
+// CK12-64-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
+// CK12-64-DAG: store i[[sz]] [[VAL:%[^,]+]], i[[sz]]* [[CBP1]]
+// CK12-64-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
+// CK12-64-DAG: [[VAL]] = load i[[sz]], i[[sz]]* [[ADDR:%.+]],
+// CK12-64-DAG: [[CADDR:%.+]] = bitcast i[[sz]]* [[ADDR]] to { float, float }*
+// CK12-64-DAG: store { float, float } {{.+}}, { float, float }* [[CADDR]],
 
-  // CK12-32-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { float, float }**
-  // CK12-32-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { float, float }**
-  // CK12-32-DAG: store { float, float }* [[DECL:%[^,]+]], { float, float }** [[CBP1]]
-  // CK12-32-DAG: store { float, float }* [[DECL]], { float, float }** [[CP1]]
+// CK12-32-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to { float, float }**
+// CK12-32-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to { float, float }**
+// CK12-32-DAG: store { float, float }* [[DECL:%[^,]+]], { float, float }** [[CBP1]]
+// CK12-32-DAG: store { float, float }* [[DECL]], { float, float }** [[CP1]]
 
-  // CK12-64: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
-  // CK12-32: call void [[KERNEL:@.+]]({ float, float }* [[DECL]])
-  #pragma omp target
+// CK12-64: call void [[KERNEL:@.+]](i[[sz]] [[VAL]])
+// CK12-32: call void [[KERNEL:@.+]]({ float, float }* [[DECL]])
+#pragma omp target
   {
     fc *= fc;
   }

diff --git a/clang/test/OpenMP/target_map_codegen_12.cpp b/clang/test/OpenMP/target_map_codegen_12.cpp
index f585f34..e468807 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_12.cpp
+++ b/clang/test/OpenMP/target_map_codegen_12.cpp
@@ -46,37 +46,43 @@
 void implicit_maps_variable_length_array (int a){
   double vla[2][a];
 
-  // CK13-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 3, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], i64* [[SGEP:%[^,]+]], {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK13-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK13-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK13-DAG: [[SGEP]] = getelementptr inbounds {{.+}}[[SS:%[^,]+]], i32 0, i32 0
+// CK13-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK13-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK13-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK13-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK13-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK13-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK13-DAG: store i64* [[SGEP:%.+]], i64** [[SARG]]
+// CK13-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK13-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK13-DAG: [[SGEP]] = getelementptr inbounds {{.+}}[[SS:%[^,]+]], i32 0, i32 0
 
-  // CK13-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK13-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK13-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[sz:64|32]]*
-  // CK13-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[sz]]*
-  // CK13-DAG: store i[[sz]] 2, i[[sz]]* [[CBP0]]
-  // CK13-DAG: store i[[sz]] 2, i[[sz]]* [[CP0]]
+// CK13-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK13-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK13-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i[[sz:64|32]]*
+// CK13-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i[[sz]]*
+// CK13-DAG: store i[[sz]] 2, i[[sz]]* [[CBP0]]
+// CK13-DAG: store i[[sz]] 2, i[[sz]]* [[CP0]]
 
-  // CK13-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
-  // CK13-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
-  // CK13-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
-  // CK13-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
-  // CK13-DAG: store i[[sz]] [[VAL:%.+]], i[[sz]]* [[CBP1]]
-  // CK13-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
+// CK13-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 1
+// CK13-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 1
+// CK13-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[sz]]*
+// CK13-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[sz]]*
+// CK13-DAG: store i[[sz]] [[VAL:%.+]], i[[sz]]* [[CBP1]]
+// CK13-DAG: store i[[sz]] [[VAL]], i[[sz]]* [[CP1]]
 
-  // CK13-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 2
-  // CK13-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 2
-  // CK13-DAG: [[S2:%.+]] = getelementptr inbounds {{.+}}[[SS]], i32 0, i32 2
-  // CK13-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to double**
-  // CK13-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double**
-  // CK13-DAG: store double* [[DECL:%.+]], double** [[CBP2]]
-  // CK13-DAG: store double* [[DECL]], double** [[CP2]]
-  // CK13-DAG: store i64 [[VALS2:%.+]], i64* [[S2]],
-  // CK13-DAG: [[VALS2]] = {{mul nuw i64 %.+, 8|sext i32 %.+ to i64}}
+// CK13-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 2
+// CK13-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 2
+// CK13-DAG: [[S2:%.+]] = getelementptr inbounds {{.+}}[[SS]], i32 0, i32 2
+// CK13-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to double**
+// CK13-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double**
+// CK13-DAG: store double* [[DECL:%.+]], double** [[CBP2]]
+// CK13-DAG: store double* [[DECL]], double** [[CP2]]
+// CK13-DAG: store i64 [[VALS2:%.+]], i64* [[S2]],
+// CK13-DAG: [[VALS2]] = {{mul nuw i64 %.+, 8|sext i32 %.+ to i64}}
 
-  // CK13: call void [[KERNEL:@.+]](i[[sz]] {{.+}}, i[[sz]] {{.+}}, double* [[DECL]])
-  #pragma omp target
+// CK13: call void [[KERNEL:@.+]](i[[sz]] {{.+}}, i[[sz]] {{.+}}, double* [[DECL]])
+#pragma omp target
   {
     vla[1][3] += 1.0;
   }

diff --git a/clang/test/OpenMP/target_map_codegen_13.cpp b/clang/test/OpenMP/target_map_codegen_13.cpp
index d59aca0..be610b4 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_13.cpp
+++ b/clang/test/OpenMP/target_map_codegen_13.cpp
@@ -66,7 +66,13 @@ void implicit_maps_class (int a){
   SSS sss(a, (double)a);
 
   // CK14: define {{.*}}void @{{.+}}foo{{.+}}([[ST]]* {{[^,]+}}, i32 {{[^,]+}})
-  // CK14-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 4, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], i64* [[SIZES:%[^,]+]], {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
+  // CK14-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK14-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK14-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK14-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK14-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK14-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK14-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
   // CK14-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
   // CK14-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
   // CK14-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]], i32 0, i32 0

diff --git a/clang/test/OpenMP/target_map_codegen_14.cpp b/clang/test/OpenMP/target_map_codegen_14.cpp
index 268445f..2f976d9 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_14.cpp
+++ b/clang/test/OpenMP/target_map_codegen_14.cpp
@@ -78,7 +78,13 @@ void implicit_maps_templated_class (int a){
   SSST<123> ssst(a, (double)a);
 
   // CK15: define {{.*}}void @{{.+}}foo{{.+}}([[ST]]* {{[^,]+}}, i32 {{[^,]+}})
-  // CK15-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 4, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], i64* [[SIZES:%[^,]+]], {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
+  // CK15-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK15-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK15-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK15-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK15-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK15-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK15-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
   // CK15-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
   // CK15-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
   // CK15-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]], i32 0, i32 0
@@ -120,7 +126,13 @@ void implicit_maps_templated_class (int a){
   ssst.foo(456);
 
   // CK15: define {{.*}}void @{{.+}}bar{{.+}}([[ST]]* {{[^,]+}}, i32 {{[^,]+}})
-  // CK15-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 4, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], i64* [[SIZES:[^,]+]], {{.+}}[[TYPES2]]{{.+}}, i8** null, i8** null)
+  // CK15-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK15-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK15-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK15-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK15-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK15-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK15-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
   // CK15-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
   // CK15-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
   // CK15-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]], i32 0, i32 0

diff --git a/clang/test/OpenMP/target_map_codegen_15.cpp b/clang/test/OpenMP/target_map_codegen_15.cpp
index 964fa1c..38787ee 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_15.cpp
+++ b/clang/test/OpenMP/target_map_codegen_15.cpp
@@ -52,7 +52,11 @@ void implicit_maps_templated_function (int a){
   int i = a;
 
   // CK16: define {{.*}}i32 @{{.+}}foo{{.+}}(i32 {{[^,]+}})
-  // CK16-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
+  // CK16-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK16-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK16-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK16-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK16-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
   // CK16-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
   // CK16-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
 

diff --git a/clang/test/OpenMP/target_map_codegen_16.cpp b/clang/test/OpenMP/target_map_codegen_16.cpp
index 68f7f71..4097e00 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_16.cpp
+++ b/clang/test/OpenMP/target_map_codegen_16.cpp
@@ -49,18 +49,22 @@ public:
 void implicit_maps_struct (int a){
   SSS s = {a, (double)a};
 
-  // CK17-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
-  // CK17-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK17-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK17-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK17-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK17-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
-  // CK17-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[ST]]**
-  // CK17-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP1]]
-  // CK17-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CP1]]
+// CK17-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK17-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK17-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK17-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK17-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK17-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK17-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK17-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK17-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK17-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
+// CK17-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to [[ST]]**
+// CK17-DAG: store [[ST]]* [[DECL:%.+]], [[ST]]** [[CBP1]]
+// CK17-DAG: store [[ST]]* [[DECL]], [[ST]]** [[CP1]]
 
-  // CK17: call void [[KERNEL:@.+]]([[ST]]* [[DECL]])
-  #pragma omp target
+// CK17: call void [[KERNEL:@.+]]([[ST]]* [[DECL]])
+#pragma omp target
   {
     s.a += 1;
     s.b += 1.0;

diff --git a/clang/test/OpenMP/target_map_codegen_17.cpp b/clang/test/OpenMP/target_map_codegen_17.cpp
index 3f774c3..7220a6b 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_17.cpp
+++ b/clang/test/OpenMP/target_map_codegen_17.cpp
@@ -51,7 +51,11 @@ void implicit_maps_template_type_capture (int a){
   int i = a;
 
   // CK18: define {{.*}}i32 @{{.+}}foo{{.+}}(i32 {{[^,]+}})
-  // CK18-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZES]]{{.+}}, {{.+}}[[TYPES]]{{.+}}, i8** null, i8** null)
+  // CK18-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK18-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK18-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK18-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK18-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
   // CK18-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
   // CK18-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
 

diff --git a/clang/test/OpenMP/target_map_codegen_18.inc b/clang/test/OpenMP/target_map_codegen_18.inc
index 9e9bbed..d03dd50 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_18.inc
+++ b/clang/test/OpenMP/target_map_codegen_18.inc
@@ -245,9 +245,13 @@ void explicit_maps_single (int ii){
   int a = ii;
 
   // Region 00
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -269,9 +273,13 @@ void explicit_maps_single (int ii){
   int b = a;
 
   // Region 00n
-  // CK19-DAG: call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00n]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00n]]{{.+}}, i8** null, i8** null, i32 1, i32 0)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -294,9 +302,13 @@ void explicit_maps_single (int ii){
   int arra[100];
 
   // Region 01
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -315,9 +327,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 02
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE02]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -337,9 +353,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 03
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE03]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -359,9 +379,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 04
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE04]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE04]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -381,9 +405,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 05
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE05]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE05]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -403,10 +431,16 @@ void explicit_maps_single (int ii){
   }
 
   // Region 06
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE06]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -429,10 +463,16 @@ void explicit_maps_single (int ii){
   }
 
   // Region 07
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE07]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -455,9 +495,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 08
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE08]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE08]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -480,9 +524,13 @@ void explicit_maps_single (int ii){
   int *pa;
 
   // Region 09
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE09]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE09]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -501,9 +549,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 10
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE10]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE10]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -525,9 +577,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 11
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE11]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE11]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -549,9 +605,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 12
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE12]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE12]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -573,10 +633,16 @@ void explicit_maps_single (int ii){
   }
 
   // Region 13
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE13]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -601,10 +667,16 @@ void explicit_maps_single (int ii){
   }
 
   // Region 14
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE14]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -629,9 +701,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 15
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE15]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE15]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -656,10 +732,16 @@ void explicit_maps_single (int ii){
   int va[ii];
 
   // Region 16
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|2}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[{{1|2}} x i{{.+}}]* [[MTYPE16]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -698,9 +780,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 17
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|2}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[{{1|2}} x i{{.+}}]* [[SIZE17]], {{.+}}getelementptr {{.+}}[{{1|2}} x i{{.+}}]* [[MTYPE17]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -735,9 +821,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 18
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|2}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[{{1|2}} x i{{.+}}]* [[SIZE18]], {{.+}}getelementptr {{.+}}[{{1|2}} x i{{.+}}]* [[MTYPE18]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -772,10 +862,16 @@ void explicit_maps_single (int ii){
   }
 
   // Region 19
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|2}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[{{1|2}} x i{{.+}}]* [[MTYPE19]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -816,9 +912,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 20
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|2}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[{{1|2}} x i{{.+}}]* [[SIZE20]], {{.+}}getelementptr {{.+}}[{{1|2}} x i{{.+}}]* [[MTYPE20]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -853,10 +953,16 @@ void explicit_maps_single (int ii){
   }
 
   // Region 21
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|2}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[{{1|2}} x i{{.+}}]* [[MTYPE21]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -897,9 +1003,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 22
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|2}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[{{1|2}} x i{{.+}}]* [[SIZE22]], {{.+}}getelementptr {{.+}}[{{1|2}} x i{{.+}}]* [[MTYPE22]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -935,9 +1045,13 @@ void explicit_maps_single (int ii){
 
   // Always.
   // Region 23
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE23]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE23]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -960,9 +1074,13 @@ void explicit_maps_single (int ii){
   int ***mptr;
 
   // Region 24
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE24]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE24]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -981,9 +1099,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 25
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE25]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE25]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1005,9 +1127,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 26
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE26]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE26]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1029,9 +1155,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 27
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE27]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE27]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1053,9 +1183,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 28
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[SIZE28]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE28]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1101,9 +1235,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 29
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[SIZE29]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE29]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1152,10 +1290,16 @@ void explicit_maps_single (int ii){
   double mva[23][ii][ii+5];
 
   // Region 30
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|4}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[{{1|4}} x i{{.+}}]* [[MTYPE30]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
   //
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1212,9 +1356,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 31
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|4}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[{{1|4}} x i{{.+}}]* [[SIZE31]], {{.+}}getelementptr {{.+}}[{{1|4}} x i{{.+}}]* [[MTYPE31]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
   //
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1276,9 +1424,13 @@ void explicit_maps_single (int ii){
   double ***mptras;
 
   // Region 32
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE32]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE32]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1297,9 +1449,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 33
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE33]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE33]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1319,9 +1475,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 34
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE34]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE34]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1341,10 +1501,16 @@ void explicit_maps_single (int ii){
   }
 
   // Region 35
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE35]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
   //
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1369,9 +1535,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 36
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE36]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE36]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1393,10 +1563,16 @@ void explicit_maps_single (int ii){
   }
 
   // Region 37
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|3}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[{{1|3}} x i{{.+}}]* [[MTYPE37]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
   //
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1442,10 +1618,16 @@ void explicit_maps_single (int ii){
   }
 
   // Region 38
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|3}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[{{1|3}} x i{{.+}}]* [[MTYPE38]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
   //
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1495,10 +1677,16 @@ void explicit_maps_single (int ii){
   }
 
   // Region 39
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|3}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[{{1|3}} x i{{.+}}]* [[MTYPE39]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
   //
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1548,10 +1736,16 @@ void explicit_maps_single (int ii){
   }
 
   // Region 40
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|3}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[{{1|3}} x i{{.+}}]* [[MTYPE40]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
   //
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1579,7 +1773,7 @@ void explicit_maps_single (int ii){
   // CK19-USE-DAG: [[SEC22]] = getelementptr {{.+}}[13 x double]* [[VAR2]], i[[Z]] [[SEC222:%[^,]+]]
   // CK19-USE-DAG: [[SEC222]] = mul nsw i[[Z]] 1, %{{[^,]+}}
 
-  // CK19-NOUSE-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-NOUSE-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
   // CK19-NOUSE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-NOUSE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
   // CK19-NOUSE-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
@@ -1602,9 +1796,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 41
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 {{1|3}}, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[{{1|3}} x i{{.+}}]* [[SIZE41]], {{.+}}getelementptr {{.+}}[{{1|3}} x i{{.+}}]* [[MTYPE41]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
   //
   // CK19-USE-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-USE-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1650,9 +1848,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 42
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[SIZE42]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE42]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1698,10 +1900,16 @@ void explicit_maps_single (int ii){
   }
 
   // Region 43 - the memory is not contiguous for this map - will map the whole last dimension.
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE43]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK19-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK19-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
   //
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -1726,9 +1934,13 @@ void explicit_maps_single (int ii){
   }
 
   // Region 44
-  // CK19-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE44]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE44]]{{.+}}, i8** null, i8** null)
-  // CK19-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK19-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK19-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK19-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK19-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK19-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK19-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK19-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK19-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
   // CK19-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
   // CK19-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0

diff --git a/clang/test/OpenMP/target_map_codegen_19.cpp b/clang/test/OpenMP/target_map_codegen_19.cpp
index 2ea0dba..2158a4a 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_19.cpp
+++ b/clang/test/OpenMP/target_map_codegen_19.cpp
@@ -57,82 +57,102 @@ void explicit_maps_references_and_function_args (int a, float b, int (&c)[10], f
   int (&cc)[10] = c;
   float *&dd = d;
 
-  // Region 00
-  // CK20-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}}, i8** null, i8** null)
-  // CK20-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK20-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK20-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK20-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK20-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK20-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK20-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
-  // CK20-DAG: store i32* [[RVAR00:%.+]], i32** [[CP0]]
-  // CK20-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
-  // CK20-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
-
-  // CK20: call void [[CALL00:@.+]](i32* {{[^,]+}})
-  #pragma omp target map(to:aa)
+// Region 00
+// CK20-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK20-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK20-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK20-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK20-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK20-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK20-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK20-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK20-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK20-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK20-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK20-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
+// CK20-DAG: store i32* [[RVAR00:%.+]], i32** [[CP0]]
+// CK20-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
+// CK20-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
+
+// CK20: call void [[CALL00:@.+]](i32* {{[^,]+}})
+#pragma omp target map(to \
+                       : aa)
   {
     aa += 1;
   }
 
-  // Region 01
-  // CK20-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null, i8** null)
-  // CK20-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK20-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK20-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK20-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK20-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [10 x i32]**
-  // CK20-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK20-DAG: store [10 x i32]* [[RVAR0:%.+]], [10 x i32]** [[CBP0]]
-  // CK20-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
-  // CK20-DAG: [[SEC0]] = getelementptr {{.*}}[10 x i32]* [[RVAR00:%.+]], i{{.+}} 0, i{{.+}} 0
-  // CK20-DAG: [[RVAR0]] = load [10 x i32]*, [10 x i32]** [[VAR0:%[^,]+]]
-  // CK20-DAG: [[RVAR00]] = load [10 x i32]*, [10 x i32]** [[VAR0]]
-
-  // CK20: call void [[CALL01:@.+]]([10 x i32]* {{[^,]+}})
-  #pragma omp target map(to:cc[:5])
+// Region 01
+// CK20-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK20-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK20-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK20-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK20-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK20-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK20-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK20-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK20-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK20-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [10 x i32]**
+// CK20-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK20-DAG: store [10 x i32]* [[RVAR0:%.+]], [10 x i32]** [[CBP0]]
+// CK20-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
+// CK20-DAG: [[SEC0]] = getelementptr {{.*}}[10 x i32]* [[RVAR00:%.+]], i{{.+}} 0, i{{.+}} 0
+// CK20-DAG: [[RVAR0]] = load [10 x i32]*, [10 x i32]** [[VAR0:%[^,]+]]
+// CK20-DAG: [[RVAR00]] = load [10 x i32]*, [10 x i32]** [[VAR0]]
+
+// CK20: call void [[CALL01:@.+]]([10 x i32]* {{[^,]+}})
+#pragma omp target map(to \
+                       : cc[:5])
   {
     cc[3] += 1;
   }
 
-  // Region 02
-  // CK20-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE02]]{{.+}}, i8** null, i8** null)
-  // CK20-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK20-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK20-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK20-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK20-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float**
-  // CK20-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
-  // CK20-DAG: store float* [[VAR0:%.+]], float** [[CBP0]]
-  // CK20-DAG: store float* [[VAR0]], float** [[CP0]]
-
-  // CK20: call void [[CALL02:@.+]](float* {{[^,]+}})
-  #pragma omp target map(from:b)
+// Region 02
+// CK20-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK20-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK20-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK20-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK20-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK20-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK20-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK20-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK20-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK20-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float**
+// CK20-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
+// CK20-DAG: store float* [[VAR0:%.+]], float** [[CBP0]]
+// CK20-DAG: store float* [[VAR0]], float** [[CP0]]
+
+// CK20: call void [[CALL02:@.+]](float* {{[^,]+}})
+#pragma omp target map(from \
+                       : b)
   {
     b += 1.0f;
   }
 
-  // Region 03
-  // CK20-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE03]]{{.+}}, i8** null, i8** null)
-  // CK20-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK20-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK20-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK20-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK20-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float**
-  // CK20-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
-  // CK20-DAG: store float* [[RVAR0:%.+]], float** [[CBP0]]
-  // CK20-DAG: store float* [[SEC0:%.+]], float** [[CP0]]
-  // CK20-DAG: [[RVAR0]] = load float*, float** [[VAR0:%[^,]+]]
-  // CK20-DAG: [[SEC0]] = getelementptr {{.*}}float* [[RVAR00:%.+]], i{{.+}} 2
-  // CK20-DAG: [[RVAR00]] = load float*, float** [[VAR0]]
-
-  // CK20: call void [[CALL03:@.+]](float* {{[^,]+}})
-  #pragma omp target map(from:d[2:3])
+// Region 03
+// CK20-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK20-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK20-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK20-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK20-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK20-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK20-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK20-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK20-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK20-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float**
+// CK20-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
+// CK20-DAG: store float* [[RVAR0:%.+]], float** [[CBP0]]
+// CK20-DAG: store float* [[SEC0:%.+]], float** [[CP0]]
+// CK20-DAG: [[RVAR0]] = load float*, float** [[VAR0:%[^,]+]]
+// CK20-DAG: [[SEC0]] = getelementptr {{.*}}float* [[RVAR00:%.+]], i{{.+}} 2
+// CK20-DAG: [[RVAR00]] = load float*, float** [[VAR0]]
+
+// CK20: call void [[CALL03:@.+]](float* {{[^,]+}})
+#pragma omp target map(from \
+                       : d [2:3])
   {
     d[2] += 1.0f;
   }

diff --git a/clang/test/OpenMP/target_map_codegen_20.cpp b/clang/test/OpenMP/target_map_codegen_20.cpp
index af7e464..aa3b170 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_20.cpp
+++ b/clang/test/OpenMP/target_map_codegen_20.cpp
@@ -105,164 +105,195 @@ struct CC {
     float la[X];
     T *lb;
 
-    // Region 00
-    // CK21-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]]{{.+}}, {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}}, i8** null, i8** null)
-    // CK21-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK21-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
-    // CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-    // CK21-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
-    // CK21-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
-    // CK21-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[VAR0:%.+]], i{{.+}} 0, i{{.+}} 0
-
-    // CK21-USE: call void [[CALL00:@.+]]([[ST]]* {{[^,]+}})
-    // CK21-NOUSE: call void [[CALL00:@.+]]()
-    #pragma omp target map(A)
+// Region 00
+// CK21-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK21-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK21-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK21-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK21-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK21-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK21-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
+// CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK21-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
+// CK21-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
+// CK21-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[VAR0:%.+]], i{{.+}} 0, i{{.+}} 0
+
+// CK21-USE: call void [[CALL00:@.+]]([[ST]]* {{[^,]+}})
+// CK21-NOUSE: call void [[CALL00:@.+]]()
+#pragma omp target map(A)
     {
 #ifdef USE
       A += 1;
 #endif
     }
 
-    // Region 01
-    // CK21-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null, i8** null)
-    // CK21-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK21-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-    // CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-    // CK21-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
-    // CK21-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
-    // CK21-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
-    // CK21-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 0
-    // CK21-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
-
-    // CK21-USE: call void [[CALL01:@.+]](i32* {{[^,]+}})
-    // CK21-NOUSE: call void [[CALL01:@.+]]()
-    #pragma omp target map(lb[:X])
+// Region 01
+// CK21-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK21-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK21-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK21-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK21-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK21-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK21-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK21-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
+// CK21-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
+// CK21-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
+// CK21-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 0
+// CK21-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
+
+// CK21-USE: call void [[CALL01:@.+]](i32* {{[^,]+}})
+// CK21-NOUSE: call void [[CALL01:@.+]]()
+#pragma omp target map(lb[:X])
     {
 #ifdef USE
       lb[4] += 1;
 #endif
     }
 
-    // Region 02
-    // CK21-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE02]]{{.+}}, i8** null, i8** null)
-    // CK21-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK21-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-    // CK21-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
-
-    // CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
-    // CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float***
-    // CK21-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
-    // CK21-DAG: store float** [[SEC0:%.+]], float*** [[CP0]]
-    // CK21-DAG: store i64 {{%.+}}, i64* [[S0]]
-    // CK21-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[VAR0]], i{{.+}} 0, i{{.+}} 2
-
-    // CK21-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
-    // CK21-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
-    // CK21-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float***
-    // CK21-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
-    // CK21-DAG: store float** [[SEC0]], float*** [[CBP1]]
-    // CK21-DAG: store float* [[SEC1:%.+]], float** [[CP1]]
-    // CK21-DAG: [[SEC1]] = getelementptr {{.*}}float* [[RVAR1:%[^,]+]], i{{.+}} 123
-    // CK21-DAG: [[RVAR1]] = load float*, float** [[SEC1_:%[^,]+]]
-    // CK21-DAG: [[SEC1_]] = getelementptr {{.*}}[[ST]]* [[VAR0]], i{{.+}} 0, i{{.+}} 2
-
-    // CK21-USE: call void [[CALL02:@.+]]([[ST]]* {{[^,]+}})
-    // CK21-NOUSE: call void [[CALL02:@.+]]()
-    #pragma omp target map(from:B[X:X+2])
+// Region 02
+// CK21-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK21-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK21-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK21-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK21-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK21-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK21-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK21-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK21-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK21-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+
+// CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
+// CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float***
+// CK21-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
+// CK21-DAG: store float** [[SEC0:%.+]], float*** [[CP0]]
+// CK21-DAG: store i64 {{%.+}}, i64* [[S0]]
+// CK21-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[VAR0]], i{{.+}} 0, i{{.+}} 2
+
+// CK21-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
+// CK21-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
+// CK21-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float***
+// CK21-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
+// CK21-DAG: store float** [[SEC0]], float*** [[CBP1]]
+// CK21-DAG: store float* [[SEC1:%.+]], float** [[CP1]]
+// CK21-DAG: [[SEC1]] = getelementptr {{.*}}float* [[RVAR1:%[^,]+]], i{{.+}} 123
+// CK21-DAG: [[RVAR1]] = load float*, float** [[SEC1_:%[^,]+]]
+// CK21-DAG: [[SEC1_]] = getelementptr {{.*}}[[ST]]* [[VAR0]], i{{.+}} 0, i{{.+}} 2
+
+// CK21-USE: call void [[CALL02:@.+]]([[ST]]* {{[^,]+}})
+// CK21-NOUSE: call void [[CALL02:@.+]]()
+#pragma omp target map(from \
+                       : B [X:X + 2])
     {
 #ifdef USE
       B[2] += 1.0f;
 #endif
     }
 
-    // Region 03
-    // CK21-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE03]]{{.+}}, i8** null, i8** null)
-    // CK21-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK21-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [123 x float]**
-    // CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [123 x float]**
-    // CK21-DAG: store [123 x float]* [[VAR0:%.+]], [123 x float]** [[CBP0]]
-    // CK21-DAG: store [123 x float]* [[VAR0]], [123 x float]** [[CP0]]
-
-    // CK21-USE: call void [[CALL03:@.+]]([123 x float]* {{[^,]+}})
-    // CK21-NOUSE: call void [[CALL03:@.+]]()
-    #pragma omp target map(from:la)
+// Region 03
+// CK21-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK21-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK21-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK21-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK21-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK21-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK21-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [123 x float]**
+// CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [123 x float]**
+// CK21-DAG: store [123 x float]* [[VAR0:%.+]], [123 x float]** [[CBP0]]
+// CK21-DAG: store [123 x float]* [[VAR0]], [123 x float]** [[CP0]]
+
+// CK21-USE: call void [[CALL03:@.+]]([123 x float]* {{[^,]+}})
+// CK21-NOUSE: call void [[CALL03:@.+]]()
+#pragma omp target map(from \
+                       : la)
     {
 #ifdef USE
       la[3] += 1.0f;
 #endif
     }
 
-    // Region 04
-    // CK21-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE04]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE04]]{{.+}}, i8** null, i8** null)
-    // CK21-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK21-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-    // CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-    // CK21-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
-    // CK21-DAG: store i32* [[VAR0]], i32** [[CP0]]
-
-    // CK21-USE: call void [[CALL04:@.+]](i32* {{[^,]+}})
-    // CK21-NOUSE: call void [[CALL04:@.+]]()
-    #pragma omp target map(from:arg)
+// Region 04
+// CK21-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK21-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK21-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK21-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK21-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK21-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK21-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK21-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
+// CK21-DAG: store i32* [[VAR0]], i32** [[CP0]]
+
+// CK21-USE: call void [[CALL04:@.+]](i32* {{[^,]+}})
+// CK21-NOUSE: call void [[CALL04:@.+]]()
+#pragma omp target map(from \
+                       : arg)
     {
 #ifdef USE
       arg +=1;
 #endif
     }
 
-    // Make sure the extra flag is passed to the second map.
-    // Region 05
-    // CK21-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE05]]{{.+}}, i8** null, i8** null)
-    // CK21-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK21-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-    // CK21-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
-
-    // CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
-    // CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
-    // CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-    // CK21-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
-    // CK21-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
-    // CK21-DAG: store i64 {{%.+}}, i64* [[S0]]
-    // CK21-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[VAR0]], i{{.+}} 0, i{{.+}} 0
-
-    // CK21-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
-    // CK21-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
-    // CK21-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
-    // CK21-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
-    // CK21-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CBP1]]
-    // CK21-DAG: store i32* [[SEC0]], i32** [[CP1]]
-
-    // CK21-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
-    // CK21-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
-    // CK21-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [[ST]]**
-    // CK21-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to i32**
-    // CK21-DAG: store [[ST]]* [[VAR2:%.+]], [[ST]]** [[CBP2]]
-    // CK21-DAG: store i32* [[SEC2:%.+]], i32** [[CP2]]
-    // CK21-DAG: [[SEC2]] = getelementptr {{.*}}[[ST]]* [[VAR2]], i{{.+}} 0, i{{.+}} 1
-
-    // CK21-USE: call void [[CALL05:@.+]]([[ST]]* {{[^,]+}})
-    // CK21-NOUSE: call void [[CALL05:@.+]]()
-    #pragma omp target map(A, A2)
+// Make sure the extra flag is passed to the second map.
+// Region 05
+// CK21-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK21-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK21-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK21-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK21-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK21-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK21-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK21-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK21-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK21-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+
+// CK21-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
+// CK21-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
+// CK21-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK21-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
+// CK21-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
+// CK21-DAG: store i64 {{%.+}}, i64* [[S0]]
+// CK21-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[VAR0]], i{{.+}} 0, i{{.+}} 0
+
+// CK21-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
+// CK21-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
+// CK21-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[ST]]**
+// CK21-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
+// CK21-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CBP1]]
+// CK21-DAG: store i32* [[SEC0]], i32** [[CP1]]
+
+// CK21-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
+// CK21-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
+// CK21-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to [[ST]]**
+// CK21-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to i32**
+// CK21-DAG: store [[ST]]* [[VAR2:%.+]], [[ST]]** [[CBP2]]
+// CK21-DAG: store i32* [[SEC2:%.+]], i32** [[CP2]]
+// CK21-DAG: [[SEC2]] = getelementptr {{.*}}[[ST]]* [[VAR2]], i{{.+}} 0, i{{.+}} 1
+
+// CK21-USE: call void [[CALL05:@.+]]([[ST]]* {{[^,]+}})
+// CK21-NOUSE: call void [[CALL05:@.+]]()
+#pragma omp target map(A, A2)
     {
 #ifdef USE
       A += 1;

diff --git a/clang/test/OpenMP/target_map_codegen_21.cpp b/clang/test/OpenMP/target_map_codegen_21.cpp
index 2828467..0048717 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_21.cpp
+++ b/clang/test/OpenMP/target_map_codegen_21.cpp
@@ -119,250 +119,310 @@ STT<int> *std;
 
 // CK22-LABEL: explicit_maps_globals{{.*}}(
 int explicit_maps_globals(void){
-  // Region 00
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK22-DAG: store i32* @a, i32** [[CBP0]]
-  // CK22-DAG: store i32* @a, i32** [[CP0]]
-
-  // CK22: call void [[CALL00:@.+]](i32* {{[^,]+}})
-  #pragma omp target map(a)
+// Region 00
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK22-DAG: store i32* @a, i32** [[CBP0]]
+// CK22-DAG: store i32* @a, i32** [[CP0]]
+
+// CK22: call void [[CALL00:@.+]](i32* {{[^,]+}})
+#pragma omp target map(a)
   { a+=1; }
 
-  // Region 01
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [100 x i32]**
-  // CK22-DAG: store [100 x i32]* @c, [100 x i32]** [[CBP0]]
-  // CK22-DAG: store [100 x i32]* @c, [100 x i32]** [[CP0]]
-
-  // CK22: call void [[CALL01:@.+]]([100 x i32]* {{[^,]+}})
-  #pragma omp target map(c)
+// Region 01
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [100 x i32]**
+// CK22-DAG: store [100 x i32]* @c, [100 x i32]** [[CBP0]]
+// CK22-DAG: store [100 x i32]* @c, [100 x i32]** [[CP0]]
+
+// CK22: call void [[CALL01:@.+]]([100 x i32]* {{[^,]+}})
+#pragma omp target map(c)
   { c[3]+=1; }
 
-  // Region 02
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE02]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
-  // CK22-DAG: store i32** @d, i32*** [[CBP0]]
-  // CK22-DAG: store i32** @d, i32*** [[CP0]]
-
-  // CK22: call void [[CALL02:@.+]](i32** {{[^,]+}})
-  #pragma omp target map(d)
+// Region 02
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
+// CK22-DAG: store i32** @d, i32*** [[CBP0]]
+// CK22-DAG: store i32** @d, i32*** [[CP0]]
+
+// CK22: call void [[CALL02:@.+]](i32** {{[^,]+}})
+#pragma omp target map(d)
   { d[3]+=1; }
 
-  // Region 03
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE03]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK22-DAG: store [100 x i32]* @c, [100 x i32]** [[CBP0]]
-  // CK22-DAG: store i32* getelementptr inbounds ([100 x i32], [100 x i32]* @c, i{{.+}} 0, i{{.+}} 1), i32** [[CP0]]
-
-  // CK22: call void [[CALL03:@.+]]([100 x i32]* {{[^,]+}})
-  #pragma omp target map(c[1:4])
+// Region 03
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x i32]**
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK22-DAG: store [100 x i32]* @c, [100 x i32]** [[CBP0]]
+// CK22-DAG: store i32* getelementptr inbounds ([100 x i32], [100 x i32]* @c, i{{.+}} 0, i{{.+}} 1), i32** [[CP0]]
+
+// CK22: call void [[CALL03:@.+]]([100 x i32]* {{[^,]+}})
+#pragma omp target map(c [1:4])
   { c[3]+=1; }
 
-  // Region 04
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE04]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE04]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK22-DAG: store i32** @d, i32*** [[CBP0]]
-  // CK22-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
-  // CK22-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 2
-  // CK22-DAG: [[RVAR00]] = load i32*, i32** @d
-
-  // CK22: call void [[CALL04:@.+]](i32* {{[^,]+}})
-  #pragma omp target map(d[2:5])
+// Region 04
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK22-DAG: store i32** @d, i32*** [[CBP0]]
+// CK22-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
+// CK22-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 2
+// CK22-DAG: [[RVAR00]] = load i32*, i32** @d
+
+// CK22: call void [[CALL04:@.+]](i32* {{[^,]+}})
+#pragma omp target map(d [2:5])
   { d[3]+=1; }
 
-  // Region 05
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE05]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE05]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
-  // CK22-DAG: store [[ST]]* @sa, [[ST]]** [[CBP0]]
-  // CK22-DAG: store [[ST]]* @sa, [[ST]]** [[CP0]]
-
-  // CK22: call void [[CALL05:@.+]]([[ST]]* {{[^,]+}})
-  #pragma omp target map(sa)
+// Region 05
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
+// CK22-DAG: store [[ST]]* @sa, [[ST]]** [[CBP0]]
+// CK22-DAG: store [[ST]]* @sa, [[ST]]** [[CP0]]
+
+// CK22: call void [[CALL05:@.+]]([[ST]]* {{[^,]+}})
+#pragma omp target map(sa)
   { sa.fa+=1; }
 
-  // Region 06
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE06]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE06]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x [[ST]]]**
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [100 x [[ST]]]**
-  // CK22-DAG: store [100 x [[ST]]]* @sc, [100 x [[ST]]]** [[CBP0]]
-  // CK22-DAG: store [100 x [[ST]]]* @sc, [100 x [[ST]]]** [[CP0]]
-
-  // CK22: call void [[CALL06:@.+]]([100 x [[ST]]]* {{[^,]+}})
-  #pragma omp target map(sc)
+// Region 06
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x [[ST]]]**
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [100 x [[ST]]]**
+// CK22-DAG: store [100 x [[ST]]]* @sc, [100 x [[ST]]]** [[CBP0]]
+// CK22-DAG: store [100 x [[ST]]]* @sc, [100 x [[ST]]]** [[CP0]]
+
+// CK22: call void [[CALL06:@.+]]([100 x [[ST]]]* {{[^,]+}})
+#pragma omp target map(sc)
   { sc[3].fa+=1; }
 
-  // Region 07
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE07]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE07]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]***
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]***
-  // CK22-DAG: store [[ST]]** @sd, [[ST]]*** [[CBP0]]
-  // CK22-DAG: store [[ST]]** @sd, [[ST]]*** [[CP0]]
-
-  // CK22: call void [[CALL07:@.+]]([[ST]]** {{[^,]+}})
-  #pragma omp target map(sd)
+// Region 07
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]***
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]***
+// CK22-DAG: store [[ST]]** @sd, [[ST]]*** [[CBP0]]
+// CK22-DAG: store [[ST]]** @sd, [[ST]]*** [[CP0]]
+
+// CK22: call void [[CALL07:@.+]]([[ST]]** {{[^,]+}})
+#pragma omp target map(sd)
   { sd[3].fa+=1; }
 
-  // Region 08
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE08]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE08]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x [[ST]]]**
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
-  // CK22-DAG: store [100 x [[ST]]]* @sc, [100 x [[ST]]]** [[CBP0]]
-  // CK22-DAG: store [[ST]]* getelementptr inbounds ([100 x [[ST]]], [100 x [[ST]]]* @sc, i{{.+}} 0, i{{.+}} 1), [[ST]]** [[CP0]]
-
-  // CK22: call void [[CALL08:@.+]]([100 x [[ST]]]* {{[^,]+}})
-  #pragma omp target map(sc[1:4])
+// Region 08
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x [[ST]]]**
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
+// CK22-DAG: store [100 x [[ST]]]* @sc, [100 x [[ST]]]** [[CBP0]]
+// CK22-DAG: store [[ST]]* getelementptr inbounds ([100 x [[ST]]], [100 x [[ST]]]* @sc, i{{.+}} 0, i{{.+}} 1), [[ST]]** [[CP0]]
+
+// CK22: call void [[CALL08:@.+]]([100 x [[ST]]]* {{[^,]+}})
+#pragma omp target map(sc [1:4])
   { sc[3].fa+=1; }
 
-  // Region 09
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE09]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE09]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]***
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
-  // CK22-DAG: store [[ST]]** @sd, [[ST]]*** [[CBP0]]
-  // CK22-DAG: store [[ST]]* [[SEC0:%.+]], [[ST]]** [[CP0]]
-  // CK22-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[RVAR00:%.+]], i{{.+}} 2
-  // CK22-DAG: [[RVAR00]] = load [[ST]]*, [[ST]]** @sd
-
-  // CK22: call void [[CALL09:@.+]]([[ST]]* {{[^,]+}})
-  #pragma omp target map(sd[2:5])
+// Region 09
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]***
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
+// CK22-DAG: store [[ST]]** @sd, [[ST]]*** [[CBP0]]
+// CK22-DAG: store [[ST]]* [[SEC0:%.+]], [[ST]]** [[CP0]]
+// CK22-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[RVAR00:%.+]], i{{.+}} 2
+// CK22-DAG: [[RVAR00]] = load [[ST]]*, [[ST]]** @sd
+
+// CK22: call void [[CALL09:@.+]]([[ST]]* {{[^,]+}})
+#pragma omp target map(sd [2:5])
   { sd[3].fa+=1; }
 
-  // Region 10
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE10]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE10]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[STT]]**
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[STT]]**
-  // CK22-DAG: store [[STT]]* @sta, [[STT]]** [[CBP0]]
-  // CK22-DAG: store [[STT]]* @sta, [[STT]]** [[CP0]]
-
-  // CK22: call void [[CALL10:@.+]]([[STT]]* {{[^,]+}})
-  #pragma omp target map(sta)
+// Region 10
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[STT]]**
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[STT]]**
+// CK22-DAG: store [[STT]]* @sta, [[STT]]** [[CBP0]]
+// CK22-DAG: store [[STT]]* @sta, [[STT]]** [[CP0]]
+
+// CK22: call void [[CALL10:@.+]]([[STT]]* {{[^,]+}})
+#pragma omp target map(sta)
   { sta.fa+=1; }
 
-  // Region 11
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE11]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE11]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x [[STT]]]**
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [100 x [[STT]]]**
-  // CK22-DAG: store [100 x [[STT]]]* @stc, [100 x [[STT]]]** [[CBP0]]
-  // CK22-DAG: store [100 x [[STT]]]* @stc, [100 x [[STT]]]** [[CP0]]
-
-  // CK22: call void [[CALL11:@.+]]([100 x [[STT]]]* {{[^,]+}})
-  #pragma omp target map(stc)
+// Region 11
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x [[STT]]]**
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [100 x [[STT]]]**
+// CK22-DAG: store [100 x [[STT]]]* @stc, [100 x [[STT]]]** [[CBP0]]
+// CK22-DAG: store [100 x [[STT]]]* @stc, [100 x [[STT]]]** [[CP0]]
+
+// CK22: call void [[CALL11:@.+]]([100 x [[STT]]]* {{[^,]+}})
+#pragma omp target map(stc)
   { stc[3].fa+=1; }
 
-  // Region 12
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE12]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE12]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[STT]]***
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[STT]]***
-  // CK22-DAG: store [[STT]]** @std, [[STT]]*** [[CBP0]]
-  // CK22-DAG: store [[STT]]** @std, [[STT]]*** [[CP0]]
-
-  // CK22: call void [[CALL12:@.+]]([[STT]]** {{[^,]+}})
-  #pragma omp target map(std)
+// Region 12
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[STT]]***
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[STT]]***
+// CK22-DAG: store [[STT]]** @std, [[STT]]*** [[CBP0]]
+// CK22-DAG: store [[STT]]** @std, [[STT]]*** [[CP0]]
+
+// CK22: call void [[CALL12:@.+]]([[STT]]** {{[^,]+}})
+#pragma omp target map(std)
   { std[3].fa+=1; }
 
-  // Region 13
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE13]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE13]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x [[STT]]]**
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[STT]]**
-  // CK22-DAG: store [100 x [[STT]]]* @stc, [100 x [[STT]]]** [[CBP0]]
-  // CK22-DAG: store [[STT]]* getelementptr inbounds ([100 x [[STT]]], [100 x [[STT]]]* @stc, i{{.+}} 0, i{{.+}} 1),  [[STT]]** [[CP0]]
-
-  // CK22: call void [[CALL13:@.+]]([100 x [[STT]]]* {{[^,]+}})
-  #pragma omp target map(stc[1:4])
+// Region 13
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x [[STT]]]**
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[STT]]**
+// CK22-DAG: store [100 x [[STT]]]* @stc, [100 x [[STT]]]** [[CBP0]]
+// CK22-DAG: store [[STT]]* getelementptr inbounds ([100 x [[STT]]], [100 x [[STT]]]* @stc, i{{.+}} 0, i{{.+}} 1),  [[STT]]** [[CP0]]
+
+// CK22: call void [[CALL13:@.+]]([100 x [[STT]]]* {{[^,]+}})
+#pragma omp target map(stc [1:4])
   { stc[3].fa+=1; }
 
-  // Region 14
-  // CK22-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE14]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE14]]{{.+}}, i8** null, i8** null)
-  // CK22-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK22-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[STT]]***
-  // CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[STT]]**
-  // CK22-DAG: store [[STT]]** @std, [[STT]]*** [[CBP0]]
-  // CK22-DAG: store [[STT]]* [[SEC0:%.+]], [[STT]]** [[CP0]]
-  // CK22-DAG: [[SEC0]] = getelementptr {{.*}}[[STT]]* [[RVAR00:%.+]], i{{.+}} 2
-  // CK22-DAG: [[RVAR00]] = load [[STT]]*, [[STT]]** @std
-
-  // CK22: call void [[CALL14:@.+]]([[STT]]* {{[^,]+}})
-  #pragma omp target map(std[2:5])
+// Region 14
+// CK22-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK22-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK22-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK22-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK22-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK22-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK22-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK22-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK22-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[STT]]***
+// CK22-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[STT]]**
+// CK22-DAG: store [[STT]]** @std, [[STT]]*** [[CBP0]]
+// CK22-DAG: store [[STT]]* [[SEC0:%.+]], [[STT]]** [[CP0]]
+// CK22-DAG: [[SEC0]] = getelementptr {{.*}}[[STT]]* [[RVAR00:%.+]], i{{.+}} 2
+// CK22-DAG: [[RVAR00]] = load [[STT]]*, [[STT]]** @std
+
+// CK22: call void [[CALL14:@.+]]([[STT]]* {{[^,]+}})
+#pragma omp target map(std [2:5])
   { std[3].fa+=1; }
 
   return 0;

diff --git a/clang/test/OpenMP/target_map_codegen_22.cpp b/clang/test/OpenMP/target_map_codegen_22.cpp
index b7dcf73..4fc42bf 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_22.cpp
+++ b/clang/test/OpenMP/target_map_codegen_22.cpp
@@ -66,125 +66,149 @@ int explicit_maps_inside_captured(int a){
   // CK23: call void @{{.*}}explicit_maps_inside_captured{{.*}}([[SA:%.+]]* {{.*}})
   // CK23: define {{.*}}explicit_maps_inside_captured{{.*}}
   [&](void){
-    // Region 00
-    // CK23-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}}, i8** null, i8** null)
-    // CK23-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK23-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-    // CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-    // CK23-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
-    // CK23-DAG: store i32* [[VAR00:%.+]], i32** [[CP0]]
-    // CK23-DAG: [[VAR0]] = load i32*, i32** [[CAP0:%[^,]+]]
-    // CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
-    // CK23-DAG: [[VAR00]] = load i32*, i32** [[CAP00:%[^,]+]]
-    // CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
-
-    // CK23: call void [[CALL00:@.+]](i32* {{[^,]+}})
-    #pragma omp target map(a)
+// Region 00
+// CK23-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK23-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK23-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK23-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK23-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK23-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
+// CK23-DAG: store i32* [[VAR00:%.+]], i32** [[CP0]]
+// CK23-DAG: [[VAR0]] = load i32*, i32** [[CAP0:%[^,]+]]
+// CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
+// CK23-DAG: [[VAR00]] = load i32*, i32** [[CAP00:%[^,]+]]
+// CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
+
+// CK23: call void [[CALL00:@.+]](i32* {{[^,]+}})
+#pragma omp target map(a)
       { a+=1; }
-    // Region 01
-    // CK23-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null, i8** null)
-    // CK23-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK23-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float**
-    // CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
-    // CK23-DAG: store float* [[VAR0:%.+]], float** [[CBP0]]
-    // CK23-DAG: store float* [[VAR00:%.+]], float** [[CP0]]
-    // CK23-DAG: [[VAR0]] = load float*, float** [[CAP0:%[^,]+]]
-    // CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
-    // CK23-DAG: [[VAR00]] = load float*, float** [[CAP00:%[^,]+]]
-    // CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
-
-    // CK23: call void [[CALL01:@.+]](float* {{[^,]+}})
-    #pragma omp target map(b)
+// Region 01
+// CK23-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK23-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK23-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK23-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK23-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float**
+// CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
+// CK23-DAG: store float* [[VAR0:%.+]], float** [[CBP0]]
+// CK23-DAG: store float* [[VAR00:%.+]], float** [[CP0]]
+// CK23-DAG: [[VAR0]] = load float*, float** [[CAP0:%[^,]+]]
+// CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
+// CK23-DAG: [[VAR00]] = load float*, float** [[CAP00:%[^,]+]]
+// CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
+
+// CK23: call void [[CALL01:@.+]](float* {{[^,]+}})
+#pragma omp target map(b)
       { b+=1; }
-    // Region 02
-    // CK23-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE02]]{{.+}}, i8** null, i8** null)
-    // CK23-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK23-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x float]**
-    // CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [100 x float]**
-    // CK23-DAG: store [100 x float]* [[VAR0:%.+]], [100 x float]** [[CBP0]]
-    // CK23-DAG: store [100 x float]* [[VAR00:%.+]], [100 x float]** [[CP0]]
-    // CK23-DAG: [[VAR0]] = load [100 x float]*, [100 x float]** [[CAP0:%[^,]+]]
-    // CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
-    // CK23-DAG: [[VAR00]] = load [100 x float]*, [100 x float]** [[CAP00:%[^,]+]]
-    // CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
-
-    // CK23: call void [[CALL02:@.+]]([100 x float]* {{[^,]+}})
-    #pragma omp target map(c)
+// Region 02
+// CK23-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK23-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK23-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK23-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK23-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x float]**
+// CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [100 x float]**
+// CK23-DAG: store [100 x float]* [[VAR0:%.+]], [100 x float]** [[CBP0]]
+// CK23-DAG: store [100 x float]* [[VAR00:%.+]], [100 x float]** [[CP0]]
+// CK23-DAG: [[VAR0]] = load [100 x float]*, [100 x float]** [[CAP0:%[^,]+]]
+// CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
+// CK23-DAG: [[VAR00]] = load [100 x float]*, [100 x float]** [[CAP00:%[^,]+]]
+// CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
+
+// CK23: call void [[CALL02:@.+]]([100 x float]* {{[^,]+}})
+#pragma omp target map(c)
       { c[3]+=1; }
 
-    // Region 03
-    // CK23-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE03]]{{.+}}, i8** null, i8** null)
-    // CK23-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK23-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float***
-    // CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float***
-    // CK23-DAG: store float** [[VAR0:%.+]], float*** [[CBP0]]
-    // CK23-DAG: store float** [[VAR00:%.+]], float*** [[CP0]]
-    // CK23-DAG: [[VAR0]] = load float**, float*** [[CAP0:%[^,]+]]
-    // CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
-    // CK23-DAG: [[VAR00]] = load float**, float*** [[CAP00:%[^,]+]]
-    // CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
-
-    // CK23: call void [[CALL03:@.+]](float** {{[^,]+}})
-    #pragma omp target map(d)
+// Region 03
+// CK23-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK23-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK23-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK23-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK23-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float***
+// CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float***
+// CK23-DAG: store float** [[VAR0:%.+]], float*** [[CBP0]]
+// CK23-DAG: store float** [[VAR00:%.+]], float*** [[CP0]]
+// CK23-DAG: [[VAR0]] = load float**, float*** [[CAP0:%[^,]+]]
+// CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
+// CK23-DAG: [[VAR00]] = load float**, float*** [[CAP00:%[^,]+]]
+// CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
+
+// CK23: call void [[CALL03:@.+]](float** {{[^,]+}})
+#pragma omp target map(d)
       { d[3]+=1; }
-    // Region 04
-    // CK23-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE04]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE04]]{{.+}}, i8** null, i8** null)
-    // CK23-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK23-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x float]**
-    // CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
-    // CK23-DAG: store [100 x float]* [[VAR0:%.+]], [100 x float]** [[CBP0]]
-    // CK23-DAG: store float* [[SEC0:%.+]], float** [[CP0]]
-    // CK23-DAG: [[SEC0]] = getelementptr {{.*}}[100 x float]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 2
-    // CK23-DAG: [[VAR0]] = load [100 x float]*, [100 x float]** [[CAP0:%[^,]+]]
-    // CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
-    // CK23-DAG: [[VAR00]] = load [100 x float]*, [100 x float]** [[CAP00:%[^,]+]]
-    // CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
-
-    // CK23: call void [[CALL04:@.+]]([100 x float]* {{[^,]+}})
-    #pragma omp target map(c[2:4])
+// Region 04
+// CK23-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK23-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK23-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK23-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK23-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [100 x float]**
+// CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
+// CK23-DAG: store [100 x float]* [[VAR0:%.+]], [100 x float]** [[CBP0]]
+// CK23-DAG: store float* [[SEC0:%.+]], float** [[CP0]]
+// CK23-DAG: [[SEC0]] = getelementptr {{.*}}[100 x float]* [[VAR00:%.+]], i{{.+}} 0, i{{.+}} 2
+// CK23-DAG: [[VAR0]] = load [100 x float]*, [100 x float]** [[CAP0:%[^,]+]]
+// CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
+// CK23-DAG: [[VAR00]] = load [100 x float]*, [100 x float]** [[CAP00:%[^,]+]]
+// CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
+
+// CK23: call void [[CALL04:@.+]]([100 x float]* {{[^,]+}})
+#pragma omp target map(c [2:4])
       { c[3]+=1; }
 
-    // Region 05
-    // CK23-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE05]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE05]]{{.+}}, i8** null, i8** null)
-    // CK23-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK23-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float**
-    // CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
-    // CK23-DAG: store float* [[RVAR0:%.+]], float** [[CBP0]]
-    // CK23-DAG: store float* [[SEC0:%.+]], float** [[CP0]]
-    // CK23-DAG: [[RVAR0]] = load float*, float** [[VAR0:%[^,]+]]
-    // CK23-DAG: [[SEC0]] = getelementptr {{.*}}float* [[RVAR00:%.+]], i{{.+}} 2
-    // CK23-DAG: [[RVAR00]] = load float*, float** [[VAR00:%[^,]+]]
-    // CK23-DAG: [[VAR0]] = load float**, float*** [[CAP0:%[^,]+]]
-    // CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
-    // CK23-DAG: [[VAR00]] = load float**, float*** [[CAP00:%[^,]+]]
-    // CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
-
-    // CK23: call void [[CALL05:@.+]](float* {{[^,]+}})
-    #pragma omp target map(d[2:4])
+// Region 05
+// CK23-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK23-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK23-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK23-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK23-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK23-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK23-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK23-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK23-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK23-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to float**
+// CK23-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to float**
+// CK23-DAG: store float* [[RVAR0:%.+]], float** [[CBP0]]
+// CK23-DAG: store float* [[SEC0:%.+]], float** [[CP0]]
+// CK23-DAG: [[RVAR0]] = load float*, float** [[VAR0:%[^,]+]]
+// CK23-DAG: [[SEC0]] = getelementptr {{.*}}float* [[RVAR00:%.+]], i{{.+}} 2
+// CK23-DAG: [[RVAR00]] = load float*, float** [[VAR00:%[^,]+]]
+// CK23-DAG: [[VAR0]] = load float**, float*** [[CAP0:%[^,]+]]
+// CK23-DAG: [[CAP0]] = getelementptr inbounds [[SA]], [[SA]]
+// CK23-DAG: [[VAR00]] = load float**, float*** [[CAP00:%[^,]+]]
+// CK23-DAG: [[CAP00]] = getelementptr inbounds [[SA]], [[SA]]
+
+// CK23: call void [[CALL05:@.+]](float* {{[^,]+}})
+#pragma omp target map(d [2:4])
       { d[3]+=1; }
   }();
   return b;

diff --git a/clang/test/OpenMP/target_map_codegen_23.cpp b/clang/test/OpenMP/target_map_codegen_23.cpp
index a9efb90..bec3e1c 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_23.cpp
+++ b/clang/test/OpenMP/target_map_codegen_23.cpp
@@ -114,9 +114,13 @@ int explicit_maps_struct_fields(int a){
   SC *p;
 
 // Region 01
-// CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]]{{.+}}, {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null)
-// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
 // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
 // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -134,9 +138,13 @@ int explicit_maps_struct_fields(int a){
 // Same thing but starting from a pointer.
 //
 // Region 13
-// CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE13]]{{.+}}, {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE13]]{{.+}}, i8** null)
-// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
 // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
 // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -154,9 +162,13 @@ int explicit_maps_struct_fields(int a){
   { p->a++; }
 
 // Region 14
-// CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE14]]{{.+}}, {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE14]]{{.+}}, i8** null)
-// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
 // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
 // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -175,9 +187,13 @@ int explicit_maps_struct_fields(int a){
   { p->a++; }
 
 // Region 15
-// CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE15]]{{.+}}, {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE15]]{{.+}}, i8** null)
-// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
 // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
 // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -197,10 +213,16 @@ int explicit_maps_struct_fields(int a){
   { p->a++; }
 
 // Region 16
-// CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE16]]{{.+}}, i8** null)
-// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-// CK24-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK24-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK24-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
 // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
 // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -228,10 +250,16 @@ int explicit_maps_struct_fields(int a){
   { p->a++; }
 
 // Region 17
-// CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE17]]{{.+}}, i8** null)
-// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-// CK24-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK24-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK24-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
 // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
 // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -262,9 +290,13 @@ int explicit_maps_struct_fields(int a){
   { p->a++; }
 
 // Region 18
-// CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE18]]{{.+}}, {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE18]]{{.+}}, i8** null)
-// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
 // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
 // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -285,10 +317,16 @@ int explicit_maps_struct_fields(int a){
   { p->a++; }
 
 // Region 19
-// CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE19]]{{.+}}, i8** null)
-// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-// CK24-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK24-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK24-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
 // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
 // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -330,10 +368,16 @@ int explicit_maps_struct_fields(int a){
   { p->a++; }
 
 // Region 20
-// CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE20]]{{.+}}, i8** null)
-// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-// CK24-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK24-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK24-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
 // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
 // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -364,10 +408,16 @@ int explicit_maps_struct_fields(int a){
   { p->a++; }
 
 // Region 21
-// CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE21]]{{.+}}, i8** null)
-// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-// CK24-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK24-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK24-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
 // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
 // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -407,10 +457,16 @@ int explicit_maps_struct_fields(int a){
   { p->a++; }
 
 // Region 22
-// CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE22]]{{.+}}, i8** null)
-// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-// CK24-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK24-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK24-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
 // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
 // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -440,10 +496,16 @@ int explicit_maps_struct_fields(int a){
   { p->a++; }
 
 // Region 23
-// CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE23]]{{.+}}, i8** null)
-// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-// CK24-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK24-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK24-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
 // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
 // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
@@ -484,10 +546,16 @@ int explicit_maps_struct_fields(int a){
   { p->a++; }
 
 // Region 24
-// CK24-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 4, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[4 x i{{.+}}]* [[MTYPE24]]{{.+}}, i8** null)
-// CK24-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK24-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-// CK24-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+// CK24-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK24-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK24-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK24-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK24-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK24-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK24-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK24-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK24-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK24-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
 // CK24-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
 // CK24-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0

diff --git a/clang/test/OpenMP/target_map_codegen_24.cpp b/clang/test/OpenMP/target_map_codegen_24.cpp
index 146530d..15249f1 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_24.cpp
+++ b/clang/test/OpenMP/target_map_codegen_24.cpp
@@ -52,41 +52,51 @@ struct CC {
   float B;
 
   int foo(T arg) {
-    // Region 00
-    // CK25-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]]{{.+}}, {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}}, i8** null)
-    // CK25-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK25-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK25-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK25-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK25-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
-    // CK25-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-    // CK25-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
-    // CK25-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
-    // CK25-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[VAR0:%.+]], i{{.+}} 0, i{{.+}} 0
-
-    // CK25: call void [[CALL00:@.+]]([[ST]]* {{[^,]+}})
-    #pragma omp target map(to:A)
+// Region 00
+// CK25-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK25-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK25-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK25-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK25-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK25-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK25-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK25-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK25-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK25-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
+// CK25-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK25-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
+// CK25-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
+// CK25-DAG: [[SEC0]] = getelementptr {{.*}}[[ST]]* [[VAR0:%.+]], i{{.+}} 0, i{{.+}} 0
+
+// CK25: call void [[CALL00:@.+]]([[ST]]* {{[^,]+}})
+#pragma omp target map(to \
+                       : A)
     {
       [&]() {
         A += 1;
       }();
     }
 
-    // Region 01
-    // CK25-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null)
-    // CK25-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK25-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-    // CK25-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK25-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK25-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-    // CK25-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-    // CK25-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
-    // CK25-DAG: store i32* [[VAR0]], i32** [[CP0]]
-
-    // CK25: call void [[CALL01:@.+]](i32* {{[^,]+}})
-    #pragma omp target map(to:arg)
+// Region 01
+// CK25-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK25-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK25-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK25-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK25-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK25-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK25-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK25-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK25-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK25-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK25-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK25-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
+// CK25-DAG: store i32* [[VAR0]], i32** [[CP0]]
+
+// CK25: call void [[CALL01:@.+]](i32* {{[^,]+}})
+#pragma omp target map(to \
+                       : arg)
     {
       [&]() {
         arg += 1;

diff --git a/clang/test/OpenMP/target_map_codegen_25.cpp b/clang/test/OpenMP/target_map_codegen_25.cpp
index b2e85d8..119bb4f 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_25.cpp
+++ b/clang/test/OpenMP/target_map_codegen_25.cpp
@@ -64,110 +64,126 @@ struct CC {
     // define {{.*}}void [[OUTCALL]]
     #pragma omp parallel firstprivate(fA,fB) private(pA,pB)
     {
-      // Region 00
-      // CK26-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE00]]{{.+}}, i8** null, i8** null)
-      // CK26-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-      // CK26-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-      // CK26-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-      // CK26-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-      // CK26-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
-      // CK26-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
-      // CK26-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
-      // CK26-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
-
-      // CK26-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
-      // CK26-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
-      // CK26-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
-      // CK26-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
-      // CK26-DAG: store i32* [[VAR1:%.+]], i32** [[CBP1]]
-      // CK26-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
-      // CK26-DAG: [[VAR1]] = load i32*, i32** [[PVT:%.+]],
-      // CK26-DAG: [[SEC1]] = load i32*, i32** [[PVT]],
-
-      // CK26: call void [[CALL00:@.+]]([[ST]]* {{[^,]+}}, i32* {{[^,]+}})
-      #pragma omp target map(fA)
+// Region 00
+// CK26-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK26-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK26-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK26-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK26-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK26-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK26-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK26-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK26-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK26-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
+// CK26-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
+// CK26-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
+// CK26-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
+
+// CK26-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
+// CK26-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
+// CK26-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
+// CK26-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
+// CK26-DAG: store i32* [[VAR1:%.+]], i32** [[CBP1]]
+// CK26-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
+// CK26-DAG: [[VAR1]] = load i32*, i32** [[PVT:%.+]],
+// CK26-DAG: [[SEC1]] = load i32*, i32** [[PVT]],
+
+// CK26: call void [[CALL00:@.+]]([[ST]]* {{[^,]+}}, i32* {{[^,]+}})
+#pragma omp target map(fA)
       {
         ++fA;
       }
 
-      // Region 01
-      // CK26-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null, i8** null)
-      // CK26-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-      // CK26-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-      // CK26-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-      // CK26-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-      // CK26-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
-      // CK26-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
-      // CK26-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
-      // CK26-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
-
-      // CK26-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
-      // CK26-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
-      // CK26-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
-      // CK26-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
-      // CK26-DAG: store float* [[VAR1:%.+]], float** [[CBP1]]
-      // CK26-DAG: store float* [[SEC1:%.+]], float** [[CP1]]
-      // CK26-DAG: [[VAR1]] = load float*, float** [[PVT:%.+]],
-      // CK26-DAG: [[SEC1]] = load float*, float** [[PVT]],
-
-      // CK26: call void [[CALL01:@.+]]([[ST]]* {{[^,]+}}, float* {{[^,]+}})
-      #pragma omp target map(fB)
+// Region 01
+// CK26-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK26-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK26-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK26-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK26-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK26-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK26-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK26-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK26-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK26-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
+// CK26-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
+// CK26-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
+// CK26-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
+
+// CK26-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
+// CK26-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
+// CK26-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
+// CK26-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
+// CK26-DAG: store float* [[VAR1:%.+]], float** [[CBP1]]
+// CK26-DAG: store float* [[SEC1:%.+]], float** [[CP1]]
+// CK26-DAG: [[VAR1]] = load float*, float** [[PVT:%.+]],
+// CK26-DAG: [[SEC1]] = load float*, float** [[PVT]],
+
+// CK26: call void [[CALL01:@.+]]([[ST]]* {{[^,]+}}, float* {{[^,]+}})
+#pragma omp target map(fB)
       {
         fB += 1.0;
       }
 
-      // Region 02
-      // CK26-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE02]]{{.+}}, i8** null, i8** null)
-      // CK26-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-      // CK26-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-      // CK26-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-      // CK26-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-      // CK26-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
-      // CK26-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
-      // CK26-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
-      // CK26-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
-
-      // CK26-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
-      // CK26-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
-      // CK26-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
-      // CK26-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
-      // CK26-DAG: store i32* [[VAR1:%.+]], i32** [[CBP1]]
-      // CK26-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
-      // CK26-DAG: [[VAR1]] = load i32*, i32** [[PVT:%.+]],
-      // CK26-DAG: [[SEC1]] = load i32*, i32** [[PVT]],
-
-      // CK26: call void [[CALL02:@.+]]([[ST]]* {{[^,]+}}, i32* {{[^,]+}})
-      #pragma omp target map(pA)
+// Region 02
+// CK26-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK26-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK26-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK26-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK26-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK26-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK26-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK26-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK26-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK26-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
+// CK26-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
+// CK26-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
+// CK26-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
+
+// CK26-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
+// CK26-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
+// CK26-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i32**
+// CK26-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i32**
+// CK26-DAG: store i32* [[VAR1:%.+]], i32** [[CBP1]]
+// CK26-DAG: store i32* [[SEC1:%.+]], i32** [[CP1]]
+// CK26-DAG: [[VAR1]] = load i32*, i32** [[PVT:%.+]],
+// CK26-DAG: [[SEC1]] = load i32*, i32** [[PVT]],
+
+// CK26: call void [[CALL02:@.+]]([[ST]]* {{[^,]+}}, i32* {{[^,]+}})
+#pragma omp target map(pA)
       {
         ++pA;
       }
 
-      // Region 01
-      // CK26-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE03]]{{.+}}, i8** null, i8** null)
-      // CK26-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-      // CK26-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-      // CK26-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-      // CK26-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-      // CK26-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
-      // CK26-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
-      // CK26-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
-      // CK26-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
-
-      // CK26-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
-      // CK26-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
-      // CK26-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
-      // CK26-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
-      // CK26-DAG: store float* [[VAR1:%.+]], float** [[CBP1]]
-      // CK26-DAG: store float* [[SEC1:%.+]], float** [[CP1]]
-      // CK26-DAG: [[VAR1]] = load float*, float** [[PVT:%.+]],
-      // CK26-DAG: [[SEC1]] = load float*, float** [[PVT]],
-
-      // CK26: call void [[CALL03:@.+]]([[ST]]* {{[^,]+}}, float* {{[^,]+}})
-      #pragma omp target map(pB)
+// Region 01
+// CK26-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK26-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK26-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK26-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK26-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK26-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK26-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK26-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK26-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK26-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[ST]]**
+// CK26-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[ST]]**
+// CK26-DAG: store [[ST]]* [[VAR0:%.+]], [[ST]]** [[CBP0]]
+// CK26-DAG: store [[ST]]* [[VAR0]], [[ST]]** [[CP0]]
+
+// CK26-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
+// CK26-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
+// CK26-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to float**
+// CK26-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to float**
+// CK26-DAG: store float* [[VAR1:%.+]], float** [[CBP1]]
+// CK26-DAG: store float* [[SEC1:%.+]], float** [[CP1]]
+// CK26-DAG: [[VAR1]] = load float*, float** [[PVT:%.+]],
+// CK26-DAG: [[SEC1]] = load float*, float** [[PVT]],
+
+// CK26: call void [[CALL03:@.+]]([[ST]]* {{[^,]+}}, float* {{[^,]+}})
+#pragma omp target map(pB)
       {
         pB += 1.0;
       }

diff --git a/clang/test/OpenMP/target_map_codegen_26.cpp b/clang/test/OpenMP/target_map_codegen_26.cpp
index 9f77428..07e62cc 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_26.cpp
+++ b/clang/test/OpenMP/target_map_codegen_26.cpp
@@ -70,83 +70,99 @@ void zero_size_section_and_private_maps (int ii){
   // Map of a pointer.
   int *pa;
 
-  // Region 00
-  // CK27-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}}, i8** null, i8** null)
-  // CK27-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK27-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK27-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
-  // CK27-DAG: store i32* [[VAR0]], i32** [[CP0]]
-
-  // CK27: call void [[CALL00:@.+]](i32* {{[^,]+}})
-  #pragma omp target
+// Region 00
+// CK27-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK27-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK27-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK27-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK27-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK27-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK27-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK27-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
+// CK27-DAG: store i32* [[VAR0]], i32** [[CP0]]
+
+// CK27: call void [[CALL00:@.+]](i32* {{[^,]+}})
+#pragma omp target
   {
     pa[50]++;
   }
 
-  // Region 01
-  // CK27-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null, i8** null)
-  // CK27-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK27-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK27-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
-  // CK27-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
-  // CK27-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
-  // CK27-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 0
-  // CK27-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
-
-  // CK27: call void [[CALL01:@.+]](i32* {{[^,]+}})
-  #pragma omp target map(pa[:0])
+// Region 01
+// CK27-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK27-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK27-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK27-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK27-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK27-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK27-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK27-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
+// CK27-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
+// CK27-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
+// CK27-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 0
+// CK27-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
+
+// CK27: call void [[CALL01:@.+]](i32* {{[^,]+}})
+#pragma omp target map(pa[:0])
   {
     pa[50]++;
   }
 
-  // Region 02
-  // CK27-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE02]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE02]]{{.+}}, i8** null, i8** null)
-  // CK27-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK27-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK27-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
-  // CK27-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
-  // CK27-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
-  // CK27-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 0
-  // CK27-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
-
-  // CK27: call void [[CALL02:@.+]](i32* {{[^,]+}})
-  #pragma omp target map(pa[0:0])
+// Region 02
+// CK27-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK27-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK27-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK27-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK27-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK27-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK27-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK27-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
+// CK27-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
+// CK27-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
+// CK27-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} 0
+// CK27-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
+
+// CK27: call void [[CALL02:@.+]](i32* {{[^,]+}})
+#pragma omp target map(pa [0:0])
   {
     pa[50]++;
   }
 
-  // Region 03
-  // CK27-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE03]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE03]]{{.+}}, i8** null, i8** null)
-  // CK27-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK27-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK27-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
-  // CK27-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
-  // CK27-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
-  // CK27-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} %{{.+}}
-  // CK27-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
-
-  // CK27: call void [[CALL03:@.+]](i32* {{[^,]+}})
-  #pragma omp target map(pa[ii:0])
+// Region 03
+// CK27-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK27-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK27-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK27-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK27-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK27-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK27-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK27-DAG: store i32* [[RVAR0:%.+]], i32** [[CBP0]]
+// CK27-DAG: store i32* [[SEC0:%.+]], i32** [[CP0]]
+// CK27-DAG: [[RVAR0]] = load i32*, i32** [[VAR0:%[^,]+]]
+// CK27-DAG: [[SEC0]] = getelementptr {{.*}}i32* [[RVAR00:%.+]], i{{.+}} %{{.+}}
+// CK27-DAG: [[RVAR00]] = load i32*, i32** [[VAR0]]
+
+// CK27: call void [[CALL03:@.+]](i32* {{[^,]+}})
+#pragma omp target map(pa [ii:0])
   {
     pa[50]++;
   }
@@ -155,82 +171,94 @@ void zero_size_section_and_private_maps (int ii){
   int pvtScl;
   int pvtArr[10];
 
-  // Region 04
-  // CK27: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 0, i8** null, i8** null, i{{64|32}}* null, i64* null, i8** null, i8** null)
-  // CK27: call void [[CALL04:@.+]]()
-  #pragma omp target private(pvtPtr)
+// Region 04
+// CK27-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK27: call void [[CALL04:@.+]]()
+#pragma omp target private(pvtPtr)
   {
     pvtPtr[5]++;
   }
 
-  // Region 05
-  // CK27-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE05]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE05]]{{.+}}, i8** null, i8** null)
-  // CK27-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK27-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK27-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
-  // CK27-DAG: store i32* [[VAR0]], i32** [[CP0]]
-
-  // CK27: call void [[CALL05:@.+]](i32* {{[^,]+}})
-  #pragma omp target firstprivate(pvtPtr)
+// Region 05
+// CK27-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK27-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK27-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK27-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK27-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK27-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK27-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK27-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
+// CK27-DAG: store i32* [[VAR0]], i32** [[CP0]]
+
+// CK27: call void [[CALL05:@.+]](i32* {{[^,]+}})
+#pragma omp target firstprivate(pvtPtr)
   {
     pvtPtr[5]++;
   }
 
-  // Region 06
-  // CK27: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 0, i8** null, i8** null, i{{64|32}}* null, i64* null, i8** null, i8** null)
-  // CK27: call void [[CALL06:@.+]]()
-  #pragma omp target private(pvtScl)
+// Region 06
+// CK27-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK27: call void [[CALL06:@.+]]()
+#pragma omp target private(pvtScl)
   {
     pvtScl++;
   }
 
-  // Region 07
-  // CK27-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{.+}}, i8* {{.+}}, i32 1, i8** [[BPGEP:%[0-9]+]], i8** [[PGEP:%[0-9]+]], {{.+}}[[SIZE07]]{{.+}}, {{.+}}[[MTYPE07]]{{.+}}, i8** null, i8** null)
-  // CK27-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
-  // CK27-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
-  // CK27-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
-  // CK27-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
-  // CK27-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[Z:64|32]]*
-  // CK27-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[Z]]*
-  // CK27-DAG: store i[[Z]] [[VAL:%.+]], i[[Z]]* [[CBP1]]
-  // CK27-DAG: store i[[Z]] [[VAL]], i[[Z]]* [[CP1]]
-  // CK27-DAG: [[VAL]] = load i[[Z]], i[[Z]]* [[ADDR:%.+]],
-  // CK27-64-DAG: [[CADDR:%.+]] = bitcast i[[Z]]* [[ADDR]] to i32*
-  // CK27-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
-
-  // CK27: call void [[CALL07:@.+]](i[[Z]] [[VAL]])
-  #pragma omp target firstprivate(pvtScl)
+// Region 07
+// CK27-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK27-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK27-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK27-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK27-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK27-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BPS:%[^,]+]], i32 0, i32 0
+// CK27-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[PS:%[^,]+]], i32 0, i32 0
+// CK27-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BPS]], i32 0, i32 0
+// CK27-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[PS]], i32 0, i32 0
+// CK27-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to i[[Z:64|32]]*
+// CK27-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to i[[Z]]*
+// CK27-DAG: store i[[Z]] [[VAL:%.+]], i[[Z]]* [[CBP1]]
+// CK27-DAG: store i[[Z]] [[VAL]], i[[Z]]* [[CP1]]
+// CK27-DAG: [[VAL]] = load i[[Z]], i[[Z]]* [[ADDR:%.+]],
+// CK27-64-DAG: [[CADDR:%.+]] = bitcast i[[Z]]* [[ADDR]] to i32*
+// CK27-64-DAG: store i32 {{.+}}, i32* [[CADDR]],
+
+// CK27: call void [[CALL07:@.+]](i[[Z]] [[VAL]])
+#pragma omp target firstprivate(pvtScl)
   {
     pvtScl++;
   }
 
-  // Region 08
-  // CK27: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 0, i8** null, i8** null, i{{64|32}}* null, i64* null, i8** null, i8** null)
-  // CK27: call void [[CALL08:@.+]]()
-  #pragma omp target private(pvtArr)
+// Region 08
+// CK27-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK27: call void [[CALL08:@.+]]()
+#pragma omp target private(pvtArr)
   {
     pvtArr[5]++;
   }
 
-  // Region 09
-  // CK27-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE09]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE09]]{{.+}}, i8** null, i8** null)
-  // CK27-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK27-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [10 x i32]**
-  // CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [10 x i32]**
-  // CK27-DAG: store [10 x i32]* [[VAR0:%.+]], [10 x i32]** [[CBP0]]
-  // CK27-DAG: store [10 x i32]* [[VAR0]], [10 x i32]** [[CP0]]
-
-  // CK27: call void [[CALL09:@.+]]([10 x i32]* {{[^,]+}})
-  #pragma omp target firstprivate(pvtArr)
+// Region 09
+// CK27-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK27-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK27-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK27-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK27-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK27-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK27-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK27-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK27-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK27-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [10 x i32]**
+// CK27-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [10 x i32]**
+// CK27-DAG: store [10 x i32]* [[VAR0:%.+]], [10 x i32]** [[CBP0]]
+// CK27-DAG: store [10 x i32]* [[VAR0]], [10 x i32]** [[CP0]]
+
+// CK27: call void [[CALL09:@.+]]([10 x i32]* {{[^,]+}})
+#pragma omp target firstprivate(pvtArr)
   {
     pvtArr[5]++;
   }

diff --git a/clang/test/OpenMP/target_map_codegen_27.cpp b/clang/test/OpenMP/target_map_codegen_27.cpp
index 7b681f0..c6e6b3a 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_27.cpp
+++ b/clang/test/OpenMP/target_map_codegen_27.cpp
@@ -45,45 +45,53 @@
 void explicit_maps_pointer_references (int *p){
   int *&a = p;
 
-  // Region 00
-  // CK28-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}}, i8** null, i8** null)
-  // CK28-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK28-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// Region 00
+// CK28-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK28-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK28-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK28-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK28-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK28-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK28-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
-  // CK28-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK28-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK28-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
-  // CK28-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
-  // CK28-DAG: store i32** [[VAR0:%.+]], i32*** [[CBP0]]
-  // CK28-DAG: store i32** [[VAR1:%.+]], i32*** [[CP0]]
-  // CK28-DAG: [[VAR0]] = load i32**, i32*** [[VAR00:%.+]],
-  // CK28-DAG: [[VAR1]] = load i32**, i32*** [[VAR11:%.+]],
+// CK28-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK28-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK28-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32***
+// CK28-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32***
+// CK28-DAG: store i32** [[VAR0:%.+]], i32*** [[CBP0]]
+// CK28-DAG: store i32** [[VAR1:%.+]], i32*** [[CP0]]
+// CK28-DAG: [[VAR0]] = load i32**, i32*** [[VAR00:%.+]],
+// CK28-DAG: [[VAR1]] = load i32**, i32*** [[VAR11:%.+]],
 
-  // CK28: call void [[CALL00:@.+]](i32** {{[^,]+}})
-  #pragma omp target map(a)
+// CK28: call void [[CALL00:@.+]](i32** {{[^,]+}})
+#pragma omp target map(a)
   {
     ++a;
   }
 
-  // Region 01
-  // CK28-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null, i8** null)
-  // CK28-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK28-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// Region 01
+// CK28-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK28-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK28-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK28-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK28-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK28-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK28-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
-  // CK28-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK28-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK28-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK28-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK28-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
-  // CK28-DAG: store i32* [[VAR1:%.+]], i32** [[CP0]]
-  // CK28-DAG: [[VAR0]] = load i32*, i32** [[VAR00:%.+]],
-  // CK28-DAG: [[VAR00]] = load i32**, i32*** [[VAR000:%.+]],
-  // CK28-DAG: [[VAR1]] = getelementptr inbounds i32, i32* [[VAR11:%.+]], i{{64|32}} 2
-  // CK28-DAG: [[VAR11]] = load i32*, i32** [[VAR111:%.+]],
-  // CK28-DAG: [[VAR111]] = load i32**, i32*** [[VAR1111:%.+]],
+// CK28-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK28-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK28-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK28-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK28-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
+// CK28-DAG: store i32* [[VAR1:%.+]], i32** [[CP0]]
+// CK28-DAG: [[VAR0]] = load i32*, i32** [[VAR00:%.+]],
+// CK28-DAG: [[VAR00]] = load i32**, i32*** [[VAR000:%.+]],
+// CK28-DAG: [[VAR1]] = getelementptr inbounds i32, i32* [[VAR11:%.+]], i{{64|32}} 2
+// CK28-DAG: [[VAR11]] = load i32*, i32** [[VAR111:%.+]],
+// CK28-DAG: [[VAR111]] = load i32**, i32*** [[VAR1111:%.+]],
 
-  // CK28: call void [[CALL01:@.+]](i32* {{[^,]+}})
-  #pragma omp target map(a[2:100])
+// CK28: call void [[CALL01:@.+]](i32* {{[^,]+}})
+#pragma omp target map(a [2:100])
   {
     ++a;
   }

diff --git a/clang/test/OpenMP/target_map_codegen_28.cpp b/clang/test/OpenMP/target_map_codegen_28.cpp
index 3d945f4..f8b4d1e 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_28.cpp
+++ b/clang/test/OpenMP/target_map_codegen_28.cpp
@@ -62,109 +62,127 @@ struct SSB{
   // CK29-LABEL: define {{.+}}foo
   void foo() {
 
-    // Region 00
-    // CK29-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i[[Z:64|32]]* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE00]]{{.+}}, i8** null, i8** null)
-
-    // CK29-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK29-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-    // CK29-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
-
-    // CK29-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK29-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK29-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
-    // CK29-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SSB]]**
-    // CK29-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SSA]]***
-    // CK29-DAG: store [[SSB]]* [[VAR0:%.+]], [[SSB]]** [[CBP0]]
-    // CK29-DAG: store [[SSA]]** [[VAR00:%.+]], [[SSA]]*** [[CP0]]
-    // CK29-DAG: store i64 %{{.+}}, i64* [[S0]]
-    // CK29-DAG: [[VAR0]] = load [[SSB]]*, [[SSB]]** %
-    // CK29-DAG: [[VAR00]] = getelementptr inbounds [[SSB]], [[SSB]]* [[VAR0]], i32 0, i32 0
-
-    // CK29-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
-    // CK29-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
-    // CK29-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to double****
-    // CK29-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double**
-    // CK29-DAG: store double*** [[VAR1:%.+]], double**** [[CBP2]]
-    // CK29-DAG: store double* [[VAR2:%.+]], double** [[CP2]]
-    // CK29-DAG: [[VAR1]] = getelementptr inbounds [[SSA]], [[SSA]]* %{{.+}}, i32 0, i32 1
-    // CK29-DAG: [[VAR2]] = getelementptr inbounds double, double* [[VAR22:%.+]], i{{.+}} 0
-    // CK29-DAG: [[VAR22]] = load double*, double** %{{.+}},
-
-    // CK29: call void [[CALL00:@.+]]([[SSB]]* {{[^,]+}})
-    #pragma omp target map(p->pr[:10])
+// Region 00
+// CK29-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK29-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK29-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK29-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK29-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK29-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK29-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+
+// CK29-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK29-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK29-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+
+// CK29-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK29-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK29-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
+// CK29-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SSB]]**
+// CK29-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SSA]]***
+// CK29-DAG: store [[SSB]]* [[VAR0:%.+]], [[SSB]]** [[CBP0]]
+// CK29-DAG: store [[SSA]]** [[VAR00:%.+]], [[SSA]]*** [[CP0]]
+// CK29-DAG: store i64 %{{.+}}, i64* [[S0]]
+// CK29-DAG: [[VAR0]] = load [[SSB]]*, [[SSB]]** %
+// CK29-DAG: [[VAR00]] = getelementptr inbounds [[SSB]], [[SSB]]* [[VAR0]], i32 0, i32 0
+
+// CK29-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
+// CK29-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
+// CK29-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to double****
+// CK29-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double**
+// CK29-DAG: store double*** [[VAR1:%.+]], double**** [[CBP2]]
+// CK29-DAG: store double* [[VAR2:%.+]], double** [[CP2]]
+// CK29-DAG: [[VAR1]] = getelementptr inbounds [[SSA]], [[SSA]]* %{{.+}}, i32 0, i32 1
+// CK29-DAG: [[VAR2]] = getelementptr inbounds double, double* [[VAR22:%.+]], i{{.+}} 0
+// CK29-DAG: [[VAR22]] = load double*, double** %{{.+}},
+
+// CK29: call void [[CALL00:@.+]]([[SSB]]* {{[^,]+}})
+#pragma omp target map(p->pr[:10])
     {
       p->pr++;
     }
 
-    // Region 01
-    // CK29-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null, i8** null)
-
-    // CK29-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK29-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-    // CK29-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
-
-    // CK29-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK29-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK29-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
-    // CK29-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SSB]]**
-    // CK29-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SSA]]****
-    // CK29-DAG: store [[SSB]]* [[VAR0]], [[SSB]]** [[CBP0]]
-    // CK29-DAG: store [[SSA]]*** [[VAR000:%.+]], [[SSA]]**** [[CP0]]
-    // CK29-DAG: store i64 %{{.+}}, i64* [[S0]]
-    // CK29-DAG: [[VAR000]] = getelementptr inbounds [[SSB]], [[SSB]]* [[VAR0]], i32 0, i32 1
-
-    // CK29-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
-    // CK29-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
-    // CK29-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[SSA]]****
-    // CK29-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double***
-    // CK29-DAG: store [[SSA]]*** [[VAR000]], [[SSA]]**** [[CBP1]]
-    // CK29-DAG: store double** [[VAR1:%.+]], double*** [[CP1]]
-    // CK29-DAG: [[VAR1]] = getelementptr inbounds [[SSA]], [[SSA]]* %{{.+}}, i32 0, i32 0
-
-    // CK29-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
-    // CK29-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
-    // CK29-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to double***
-    // CK29-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double**
-    // CK29-DAG: store double** [[VAR1]], double*** [[CBP2]]
-    // CK29-DAG: store double* [[VAR2:%.+]], double** [[CP2]]
-    // CK29-DAG: [[VAR2]] = getelementptr inbounds double, double* [[VAR22:%.+]], i{{.+}} 0
-    // CK29-DAG: [[VAR22]] = load double*, double** %{{.+}},
-
-    // CK29: call void [[CALL00:@.+]]([[SSB]]* {{[^,]+}})
-    #pragma omp target map(pr->p[:10])
+// Region 01
+// CK29-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK29-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK29-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK29-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK29-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK29-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK29-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+
+// CK29-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK29-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK29-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+
+// CK29-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK29-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK29-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
+// CK29-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SSB]]**
+// CK29-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SSA]]****
+// CK29-DAG: store [[SSB]]* [[VAR0]], [[SSB]]** [[CBP0]]
+// CK29-DAG: store [[SSA]]*** [[VAR000:%.+]], [[SSA]]**** [[CP0]]
+// CK29-DAG: store i64 %{{.+}}, i64* [[S0]]
+// CK29-DAG: [[VAR000]] = getelementptr inbounds [[SSB]], [[SSB]]* [[VAR0]], i32 0, i32 1
+
+// CK29-DAG: [[BP1:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
+// CK29-DAG: [[P1:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
+// CK29-DAG: [[CBP1:%.+]] = bitcast i8** [[BP1]] to [[SSA]]****
+// CK29-DAG: [[CP1:%.+]] = bitcast i8** [[P1]] to double***
+// CK29-DAG: store [[SSA]]*** [[VAR000]], [[SSA]]**** [[CBP1]]
+// CK29-DAG: store double** [[VAR1:%.+]], double*** [[CP1]]
+// CK29-DAG: [[VAR1]] = getelementptr inbounds [[SSA]], [[SSA]]* %{{.+}}, i32 0, i32 0
+
+// CK29-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 2
+// CK29-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 2
+// CK29-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to double***
+// CK29-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double**
+// CK29-DAG: store double** [[VAR1]], double*** [[CBP2]]
+// CK29-DAG: store double* [[VAR2:%.+]], double** [[CP2]]
+// CK29-DAG: [[VAR2]] = getelementptr inbounds double, double* [[VAR22:%.+]], i{{.+}} 0
+// CK29-DAG: [[VAR22]] = load double*, double** %{{.+}},
+
+// CK29: call void [[CALL00:@.+]]([[SSB]]* {{[^,]+}})
+#pragma omp target map(pr->p[:10])
     {
       pr->p++;
     }
 
-    // Region 02
-    // CK29-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE02]]{{.+}}, i8** null, i8** null)
-
-    // CK29-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-    // CK29-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-    // CK29-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
-
-    // CK29-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-    // CK29-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-    // CK29-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
-    // CK29-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SSB]]**
-    // CK29-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SSA]]****
-    // CK29-DAG: store [[SSB]]* [[VAR0]], [[SSB]]** [[CBP0]]
-    // CK29-DAG: store [[SSA]]*** [[VAR000:%.+]], [[SSA]]**** [[CP0]]
-    // CK29-DAG: store i64 %{{.+}}, i64* [[S0]]
-    // CK29-DAG: [[VAR000]] = getelementptr inbounds [[SSB]], [[SSB]]* [[VAR0]], i32 0, i32 1
-
-    // CK29-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
-    // CK29-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
-    // CK29-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to double****
-    // CK29-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double**
-    // CK29-DAG: store double*** [[VAR1:%.+]], double**** [[CBP2]]
-    // CK29-DAG: store double* [[VAR2:%.+]], double** [[CP2]]
-    // CK29-DAG: [[VAR1]] = getelementptr inbounds [[SSA]], [[SSA]]* %{{.+}}, i32 0, i32 1
-    // CK29-DAG: [[VAR2]] = getelementptr inbounds double, double* [[VAR22:%.+]], i{{.+}} 0
-    // CK29-DAG: [[VAR22]] = load double*, double** %{{.+}},
-
-    // CK29: call void [[CALL00:@.+]]([[SSB]]* {{[^,]+}})
-    #pragma omp target map(pr->pr[:10])
+// Region 02
+// CK29-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK29-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK29-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK29-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK29-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK29-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK29-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+
+// CK29-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK29-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK29-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+
+// CK29-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK29-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK29-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
+// CK29-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to [[SSB]]**
+// CK29-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to [[SSA]]****
+// CK29-DAG: store [[SSB]]* [[VAR0]], [[SSB]]** [[CBP0]]
+// CK29-DAG: store [[SSA]]*** [[VAR000:%.+]], [[SSA]]**** [[CP0]]
+// CK29-DAG: store i64 %{{.+}}, i64* [[S0]]
+// CK29-DAG: [[VAR000]] = getelementptr inbounds [[SSB]], [[SSB]]* [[VAR0]], i32 0, i32 1
+
+// CK29-DAG: [[BP2:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 1
+// CK29-DAG: [[P2:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 1
+// CK29-DAG: [[CBP2:%.+]] = bitcast i8** [[BP2]] to double****
+// CK29-DAG: [[CP2:%.+]] = bitcast i8** [[P2]] to double**
+// CK29-DAG: store double*** [[VAR1:%.+]], double**** [[CBP2]]
+// CK29-DAG: store double* [[VAR2:%.+]], double** [[CP2]]
+// CK29-DAG: [[VAR1]] = getelementptr inbounds [[SSA]], [[SSA]]* %{{.+}}, i32 0, i32 1
+// CK29-DAG: [[VAR2]] = getelementptr inbounds double, double* [[VAR22:%.+]], i{{.+}} 0
+// CK29-DAG: [[VAR22]] = load double*, double** %{{.+}},
+
+// CK29: call void [[CALL00:@.+]]([[SSB]]* {{[^,]+}})
+#pragma omp target map(pr->pr[:10])
     {
       pr->pr++;
     }

diff --git a/clang/test/OpenMP/target_map_codegen_29.cpp b/clang/test/OpenMP/target_map_codegen_29.cpp
index 22555f2..34677d2 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_29.cpp
+++ b/clang/test/OpenMP/target_map_codegen_29.cpp
@@ -56,10 +56,16 @@ typedef struct StructWithPtrTag : public Base {
   int *ptr1;
 } StructWithPtr;
 
-// CK30-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @.__omp_offloading_{{.*}}map_with_deep_copy{{.*}}_l{{[0-9]+}}.region_id, i32 4, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* [[MTYPE00]], i32 0, i32 0), i8** null, i8** null)
-// CK30-DAG: [[GEPS]] = getelementptr inbounds [4 x i{{64|32}}], [4 x i64]* [[SIZES:%.+]], i32 0, i32 0
-// CK30-DAG: [[GEPP]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[PTRS:%.+]], i32 0, i32 0
-// CK30-DAG: [[GEPBP]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[BASES:%.+]], i32 0, i32 0
+// CK30-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK30-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK30-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK30-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK30-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK30-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK30-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK30-DAG: [[SIZES]] = getelementptr inbounds [4 x i{{64|32}}], [4 x i64]* [[SIZES:%.+]], i32 0, i32 0
+// CK30-DAG: [[PGEP]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[PTRS:%.+]], i32 0, i32 0
+// CK30-DAG: [[BPGEP]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[BASES:%.+]], i32 0, i32 0
 
 // CK30-DAG: [[BASE_PTR:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[BASES]], i32 0, i32 0
 // CK30-DAG: [[BC:%.+]] = bitcast i8** [[BASE_PTR]] to [[STRUCT]]**

diff --git a/clang/test/OpenMP/target_map_codegen_30.cpp b/clang/test/OpenMP/target_map_codegen_30.cpp
index b1d23b4..bcaaada 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_30.cpp
+++ b/clang/test/OpenMP/target_map_codegen_30.cpp
@@ -46,40 +46,50 @@ void explicit_maps_single (int ii){
   // Map of a scalar.
   int a = ii;
 
-  // Close.
-  // Region 00
-  // CK31-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE00]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE00]]{{.+}}, i8** null, i8** null)
-  // CK31-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK31-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// Close.
+// Region 00
+// CK31-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK31-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK31-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK31-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK31-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK31-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK31-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
-  // CK31-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK31-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK31-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK31-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK31-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
-  // CK31-DAG: store i32* [[VAR0]], i32** [[CP0]]
+// CK31-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK31-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK31-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK31-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK31-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
+// CK31-DAG: store i32* [[VAR0]], i32** [[CP0]]
 
-  // CK31: call void [[CALL00:@.+]](i32* {{[^,]+}})
-  #pragma omp target map(close, tofrom: a)
+// CK31: call void [[CALL00:@.+]](i32* {{[^,]+}})
+#pragma omp target map(close, tofrom \
+                       : a)
   {
    a++;
   }
 
-  // Always Close.
-  // Region 01
-  // CK31-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}}, i8** null, i8** null)
-  // CK31-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK31-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// Always Close.
+// Region 01
+// CK31-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK31-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK31-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK31-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK31-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK31-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK31-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
-  // CK31-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK31-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK31-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK31-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK31-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
-  // CK31-DAG: store i32* [[VAR0]], i32** [[CP0]]
+// CK31-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK31-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK31-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK31-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK31-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
+// CK31-DAG: store i32* [[VAR0]], i32** [[CP0]]
 
-  // CK31: call void [[CALL01:@.+]](i32* {{[^,]+}})
-  #pragma omp target map(always close tofrom: a)
+// CK31: call void [[CALL01:@.+]](i32* {{[^,]+}})
+#pragma omp target map(always close tofrom \
+                       : a)
   {
    a++;
   }

diff --git a/clang/test/OpenMP/target_map_codegen_31.cpp b/clang/test/OpenMP/target_map_codegen_31.cpp
index e935bf7..b2a8b29 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_31.cpp
+++ b/clang/test/OpenMP/target_map_codegen_31.cpp
@@ -85,10 +85,16 @@ void explicit_maps_single (int ii){
 // CK31A: [[ST1_I:%.+]] = getelementptr inbounds [[ST]], [[ST]]* [[ST1]], i{{.+}} 0, i{{.+}} 0
 // CK31A: [[ST2_I:%.+]] = getelementptr inbounds [[ST]], [[ST]]* [[ST2]], i{{.+}} 0, i{{.+}} 0
 // CK31A: [[ST2_J:%.+]] = getelementptr inbounds [[ST]], [[ST]]* [[ST2]], i{{.+}} 0, i{{.+}} 1
-// CK31A-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 7, i8** [[GEPBP:%[0-9]+]], i8** [[GEPP:%[0-9]+]], i64* [[GEPS:%.+]], i64* getelementptr {{.+}}[7 x i{{.+}}]* [[MTYPE00]]{{.+}})
-// CK31A-DAG: [[GEPS]] = getelementptr inbounds [7 x i64], [7 x i64]* [[S:%.+]], i{{.+}} 0, i{{.+}} 0
-// CK31A-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK31A-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK31A-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK31A-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK31A-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK31A-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK31A-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK31A-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK31A-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK31A-DAG: [[SIZES]] = getelementptr inbounds [7 x i64], [7 x i64]* [[S:%.+]], i{{.+}} 0, i{{.+}} 0
+// CK31A-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK31A-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
 // st1
 // CK31A-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
@@ -166,22 +172,27 @@ void explicit_maps_single (int ii){
 #endif
   }
 
-  // Always Close Present.
-  // Region 01
-  // CK31A-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[SIZE01]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE01]]{{.+}})
-  // CK31A-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK31A-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-
-  // CK31A-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK31A-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK31A-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
-  // CK31A-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
-  // CK31A-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
-  // CK31A-DAG: store i32* [[VAR0]], i32** [[CP0]]
-
-  // CK31A-USE: call void [[CALL01:@.+]](i32* {{[^,]+}})
-  // CK31A-NOUSE: call void [[CALL01:@.+]]()
-  #pragma omp target map(always close present tofrom: a)
+// Always Close Present.
+// Region 01
+// CK31A-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK31A-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK31A-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK31A-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK31A-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK31A-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK31A-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+
+// CK31A-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK31A-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK31A-DAG: [[CBP0:%.+]] = bitcast i8** [[BP0]] to i32**
+// CK31A-DAG: [[CP0:%.+]] = bitcast i8** [[P0]] to i32**
+// CK31A-DAG: store i32* [[VAR0:%.+]], i32** [[CBP0]]
+// CK31A-DAG: store i32* [[VAR0]], i32** [[CP0]]
+
+// CK31A-USE: call void [[CALL01:@.+]](i32* {{[^,]+}})
+// CK31A-NOUSE: call void [[CALL01:@.+]]()
+#pragma omp target map(always close present tofrom \
+                       : a)
   {
 #ifdef USE
     a++;

diff --git a/clang/test/OpenMP/target_map_codegen_32.cpp b/clang/test/OpenMP/target_map_codegen_32.cpp
index c5c039f..be8c719 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_32.cpp
+++ b/clang/test/OpenMP/target_map_codegen_32.cpp
@@ -56,10 +56,16 @@ struct ST {
 // Region 00
 // CK31B: [[J:%.+]] = getelementptr inbounds [[ST]], [[ST]]* [[THIS:%.+]], i{{.+}} 0, i{{.+}} 1
 // CK31B: [[I:%.+]] = getelementptr inbounds [[ST]], [[ST]]* [[THIS]], i{{.+}} 0, i{{.+}} 0
-// CK31B-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 {{[^,]+}}, i8* {{[^,]+}}, i32 3, i8** [[GEPBP:%[0-9]+]], i8** [[GEPP:%[0-9]+]], i64* [[GEPS:%.+]], i64* getelementptr {{.+}}[3 x i{{.+}}]* [[MTYPE00]]{{.+}})
-// CK31B-DAG: [[GEPS]] = getelementptr inbounds [3 x i64], [3 x i64]* [[S:%.+]], i{{.+}} 0, i{{.+}} 0
-// CK31B-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-// CK31B-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK31B-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK31B-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK31B-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK31B-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK31B-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK31B-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK31B-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK31B-DAG: [[SIZES]] = getelementptr inbounds [3 x i64], [3 x i64]* [[S:%.+]], i{{.+}} 0, i{{.+}} 0
+// CK31B-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK31B-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
 
 // this
 // CK31B-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0

diff --git a/clang/test/OpenMP/target_map_codegen_33.cpp b/clang/test/OpenMP/target_map_codegen_33.cpp
index b930355..431da04 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_33.cpp
+++ b/clang/test/OpenMP/target_map_codegen_33.cpp
@@ -24,56 +24,70 @@
 
 void array_shaping(float *f, int sa) {
 
-  // CK32-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE_TO]]{{.+}}, i8** null, i8** null)
-  // CK32-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK32-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK32-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+// CK32-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK32-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK32-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK32-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK32-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK32-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK32-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK32-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK32-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK32-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
-  // CK32-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK32-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK32-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
+// CK32-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK32-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK32-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
 
-  // CK32-DAG: [[BPC0:%.+]] = bitcast i8** [[BP0]] to float**
-  // CK32-DAG: [[PC0:%.+]] = bitcast i8** [[P0]] to float**
+// CK32-DAG: [[BPC0:%.+]] = bitcast i8** [[BP0]] to float**
+// CK32-DAG: [[PC0:%.+]] = bitcast i8** [[P0]] to float**
 
-  // CK32-DAG: store float* [[F1:%.+]], float** [[BPC0]],
-  // CK32-DAG: store float* [[F2:%.+]], float** [[PC0]],
-  // CK32-DAG: store i64 [[SIZE:%.+]], i64* [[S0]],
+// CK32-DAG: store float* [[F1:%.+]], float** [[BPC0]],
+// CK32-DAG: store float* [[F2:%.+]], float** [[PC0]],
+// CK32-DAG: store i64 [[SIZE:%.+]], i64* [[S0]],
 
-  // CK32-DAG: [[F1]] = load float*, float** [[F_ADDR:%.+]],
-  // CK32-DAG: [[F2]] = load float*, float** [[F_ADDR]],
-  // CK32-64-DAG: [[SIZE]] = mul nuw i64 [[SZ1:%.+]], 4
-  // CK32-64-DAG: [[SZ1]] = mul nuw i64 12, %{{.+}}
-  // CK32-32-DAG: [[SIZE]] = sext i32 [[SZ1:%.+]] to i64
-  // CK32-32-DAG: [[SZ1]] = mul nuw i32 [[SZ2:%.+]], 4
-  // CK32-32-DAG: [[SZ2]] = mul nuw i32 12, %{{.+}}
-  #pragma omp target map(to:([3][sa][4])f)
+// CK32-DAG: [[F1]] = load float*, float** [[F_ADDR:%.+]],
+// CK32-DAG: [[F2]] = load float*, float** [[F_ADDR]],
+// CK32-64-DAG: [[SIZE]] = mul nuw i64 [[SZ1:%.+]], 4
+// CK32-64-DAG: [[SZ1]] = mul nuw i64 12, %{{.+}}
+// CK32-32-DAG: [[SIZE]] = sext i32 [[SZ1:%.+]] to i64
+// CK32-32-DAG: [[SZ1]] = mul nuw i32 [[SZ2:%.+]], 4
+// CK32-32-DAG: [[SZ2]] = mul nuw i32 12, %{{.+}}
+#pragma omp target map(to \
+                       : ([3][sa][4])f)
   f[0] = 1;
   sa = 1;
-  // CK32-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[1 x i{{.+}}]* [[MTYPE_FROM]]{{.+}}, i8** null, i8** null)
-  // CK32-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK32-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK32-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+// CK32-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CK32-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CK32-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+// CK32-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CK32-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+// CK32-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CK32-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+// CK32-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+// CK32-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+// CK32-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
-  // CK32-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
-  // CK32-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
-  // CK32-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
+// CK32-DAG: [[BP0:%.+]] = getelementptr inbounds {{.+}}[[BP]], i{{.+}} 0, i{{.+}} 0
+// CK32-DAG: [[P0:%.+]] = getelementptr inbounds {{.+}}[[P]], i{{.+}} 0, i{{.+}} 0
+// CK32-DAG: [[S0:%.+]] = getelementptr inbounds {{.+}}[[S]], i{{.+}} 0, i{{.+}} 0
 
-  // CK32-DAG: [[BPC0:%.+]] = bitcast i8** [[BP0]] to float**
-  // CK32-DAG: [[PC0:%.+]] = bitcast i8** [[P0]] to float**
+// CK32-DAG: [[BPC0:%.+]] = bitcast i8** [[BP0]] to float**
+// CK32-DAG: [[PC0:%.+]] = bitcast i8** [[P0]] to float**
 
-  // CK32-DAG: store float* [[F1:%.+]], float** [[BPC0]],
-  // CK32-DAG: store float* [[F2:%.+]], float** [[PC0]],
-  // CK32-DAG: store i64 [[SIZE:%.+]], i64* [[S0]],
+// CK32-DAG: store float* [[F1:%.+]], float** [[BPC0]],
+// CK32-DAG: store float* [[F2:%.+]], float** [[PC0]],
+// CK32-DAG: store i64 [[SIZE:%.+]], i64* [[S0]],
 
-  // CK32-DAG: [[F1]] = load float*, float** [[F_ADDR:%.+]],
-  // CK32-DAG: [[F2]] = load float*, float** [[F_ADDR]],
-  // CK32-64-DAG: [[SIZE]] = mul nuw i64 [[SZ1:%.+]], 5
-  // CK32-64-DAG: [[SZ1]] = mul nuw i64 4, %{{.+}}
-  // CK32-32-DAG: [[SIZE]] = sext i32 [[SZ1:%.+]] to i64
-  // CK32-32-DAG: [[SZ1]] = mul nuw i32 [[SZ2:%.+]], 5
-  // CK32-32-DAG: [[SZ2]] = mul nuw i32 4, %{{.+}}
-  #pragma omp target map(from: ([sa][5])f)
+// CK32-DAG: [[F1]] = load float*, float** [[F_ADDR:%.+]],
+// CK32-DAG: [[F2]] = load float*, float** [[F_ADDR]],
+// CK32-64-DAG: [[SIZE]] = mul nuw i64 [[SZ1:%.+]], 5
+// CK32-64-DAG: [[SZ1]] = mul nuw i64 4, %{{.+}}
+// CK32-32-DAG: [[SIZE]] = sext i32 [[SZ1:%.+]] to i64
+// CK32-32-DAG: [[SZ1]] = mul nuw i32 [[SZ2:%.+]], 5
+// CK32-32-DAG: [[SZ2]] = mul nuw i32 4, %{{.+}}
+#pragma omp target map(from \
+                       : ([sa][5])f)
   f[0] = 1;
 }
 

diff --git a/clang/test/OpenMP/target_map_codegen_34.cpp b/clang/test/OpenMP/target_map_codegen_34.cpp
index 94b86d6..737cc73 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_34.cpp
+++ b/clang/test/OpenMP/target_map_codegen_34.cpp
@@ -49,11 +49,19 @@ public:
 void default_mapper() {
   S s;
 
-  // CK34-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 4, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[4 x i{{.+}}]* [[MTYPE_TO]]{{.+}}, i8** null, i8** [[GEPMF:%.+]])
-  // CK34-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK34-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK34-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
-  // CK34-DAG: [[GEPMF]] = bitcast [4 x i8*]* [[MF:%.+]] to i8**
+  // CK34-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK34-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK34-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK34-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK34-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK34-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK34-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK34-DAG: [[MARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 7
+  // CK34-DAG: store i8** [[MFGEP:%.+]], i8*** [[MARG]]
+  // CK34-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK34-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK34-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK34-DAG: [[MFGEP]] = bitcast [4 x i8*]* [[MF:%.+]] to i8**
 
   // pass TARGET_PARAM {&s, &s, ((void*)(&s+1)-(void*)&s)}
 
@@ -150,11 +158,19 @@ void default_mapper() {
 
   // CK34 : call void
 
-  // CK34-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 4, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[4 x i{{.+}}]* [[MTYPE_FROM]]{{.+}}, i8** null, i8** [[GEPMF:%.+]])
-  // CK34-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK34-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK34-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
-  // CK34-DAG: [[GEPMF]] = bitcast [4 x i8*]* [[MF:%.+]] to i8**
+  // CK34-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK34-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK34-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK34-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK34-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK34-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK34-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK34-DAG: [[MARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 7
+  // CK34-DAG: store i8** [[MFGEP:%.+]], i8*** [[MARG]]
+  // CK34-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK34-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK34-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK34-DAG: [[MFGEP]] = bitcast [4 x i8*]* [[MF:%.+]] to i8**
 
   // pass TARGET_PARAM {&s, &s, ((void*)(&s+1)-(void*)&s)}
 

diff --git a/clang/test/OpenMP/target_map_codegen_35.cpp b/clang/test/OpenMP/target_map_codegen_35.cpp
index 9c09999..c46a3c5 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_35.cpp
+++ b/clang/test/OpenMP/target_map_codegen_35.cpp
@@ -41,10 +41,16 @@ void ref_map() {
   double b;
   S s(b);
 
-  // CK35-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 4, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[4 x i{{.+}}]* [[MTYPE_TO]]{{.+}}, i8** null, i8** null)
-  // CK35-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK35-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK35-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK35-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK35-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK35-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK35-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK35-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK35-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK35-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK35-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK35-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK35-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
   // pass TARGET_PARAM {&s, &s, ((void*)(&s+1)-(void*)&s)}
 
@@ -132,10 +138,16 @@ void ref_map() {
 
   // CK35 : call void
 
-  // CK35-DAG: call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 2, i8** [[GEPBP:%.+]], i8** [[GEPP:%.+]], i64* [[GEPS:%.+]], {{.+}}getelementptr {{.+}}[2 x i{{.+}}]* [[MTYPE_FROM]]{{.+}}, i8** null, i8** null)
-  // CK35-DAG: [[GEPBP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
-  // CK35-DAG: [[GEPP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
-  // CK35-DAG: [[GEPS]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
+  // CK35-DAG: call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CK35-DAG: [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CK35-DAG: store i8** [[BPGEP:%.+]], i8*** [[BPARG]]
+  // CK35-DAG: [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CK35-DAG: store i8** [[PGEP:%.+]], i8*** [[PARG]]
+  // CK35-DAG: [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+  // CK35-DAG: store i64* [[SIZES:%.+]], i64** [[SARG]]
+  // CK35-DAG: [[BPGEP]] = getelementptr inbounds {{.+}}[[BP:%[^,]+]]
+  // CK35-DAG: [[PGEP]] = getelementptr inbounds {{.+}}[[P:%[^,]+]]
+  // CK35-DAG: [[SIZES]] = getelementptr inbounds {{.+}}[[S:%[^,]+]]
 
   // pass TARGET_PARAM {&s, &s.b, ((void*)(&s.b+1)-(void*)&s.b)}
 

diff --git a/clang/test/OpenMP/target_map_codegen_hold.cpp b/clang/test/OpenMP/target_map_codegen_hold.cpp
index dff953a..128ff31 100644 (file)
--- a/clang/test/OpenMP/target_map_codegen_hold.cpp
+++ b/clang/test/OpenMP/target_map_codegen_hold.cpp
@@ -216,30 +216,64 @@ struct ST {
 // CHECK-USE-PPC64LE-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [7 x i8*], [7 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK-USE-PPC64LE-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [7 x i8*], [7 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK-USE-PPC64LE-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [7 x i64], [7 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK-USE-PPC64LE-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l630.region_id, i32 7, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([7 x i64], [7 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null)
-// CHECK-USE-PPC64LE-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
-// CHECK-USE-PPC64LE-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK-USE-PPC64LE-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK-USE-PPC64LE-NEXT:    store i32 1, i32* [[TMP56]], align 4
+// CHECK-USE-PPC64LE-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK-USE-PPC64LE-NEXT:    store i32 7, i32* [[TMP57]], align 4
+// CHECK-USE-PPC64LE-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK-USE-PPC64LE-NEXT:    store i8** [[TMP53]], i8*** [[TMP58]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK-USE-PPC64LE-NEXT:    store i8** [[TMP54]], i8*** [[TMP59]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK-USE-PPC64LE-NEXT:    store i64* [[TMP55]], i64** [[TMP60]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK-USE-PPC64LE-NEXT:    store i64* getelementptr inbounds ([7 x i64], [7 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP61]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK-USE-PPC64LE-NEXT:    store i8** null, i8*** [[TMP62]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK-USE-PPC64LE-NEXT:    store i8** null, i8*** [[TMP63]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP64:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l766.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK-USE-PPC64LE-NEXT:    [[TMP65:%.*]] = icmp ne i32 [[TMP64]], 0
+// CHECK-USE-PPC64LE-NEXT:    br i1 [[TMP65]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK-USE-PPC64LE:       omp_offload.failed:
-// CHECK-USE-PPC64LE-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l630(%struct.ST* [[ST1]], i32* [[A]], %struct.ST* [[ST2]]) #[[ATTR3:[0-9]+]]
+// CHECK-USE-PPC64LE-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l766(%struct.ST* [[ST1]], i32* [[A]], %struct.ST* [[ST2]]) #[[ATTR3:[0-9]+]]
 // CHECK-USE-PPC64LE-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK-USE-PPC64LE:       omp_offload.cont:
-// CHECK-USE-PPC64LE-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK-USE-PPC64LE-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to i32**
-// CHECK-USE-PPC64LE-NEXT:    store i32* [[A]], i32** [[TMP59]], align 8
-// CHECK-USE-PPC64LE-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK-USE-PPC64LE-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to i32**
-// CHECK-USE-PPC64LE-NEXT:    store i32* [[A]], i32** [[TMP61]], align 8
-// CHECK-USE-PPC64LE-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK-USE-PPC64LE-NEXT:    store i8* null, i8** [[TMP62]], align 8
-// CHECK-USE-PPC64LE-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK-USE-PPC64LE-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK-USE-PPC64LE-NEXT:    [[TMP65:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l644.region_id, i32 1, i8** [[TMP63]], i8** [[TMP64]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.1, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.2, i32 0, i32 0), i8** null, i8** null)
-// CHECK-USE-PPC64LE-NEXT:    [[TMP66:%.*]] = icmp ne i32 [[TMP65]], 0
-// CHECK-USE-PPC64LE-NEXT:    br i1 [[TMP66]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK-USE-PPC64LE:       omp_offload.failed6:
-// CHECK-USE-PPC64LE-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l644(i32* [[A]]) #[[ATTR3]]
-// CHECK-USE-PPC64LE-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK-USE-PPC64LE:       omp_offload.cont7:
+// CHECK-USE-PPC64LE-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK-USE-PPC64LE-NEXT:    [[TMP67:%.*]] = bitcast i8** [[TMP66]] to i32**
+// CHECK-USE-PPC64LE-NEXT:    store i32* [[A]], i32** [[TMP67]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK-USE-PPC64LE-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i32**
+// CHECK-USE-PPC64LE-NEXT:    store i32* [[A]], i32** [[TMP69]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK-USE-PPC64LE-NEXT:    store i8* null, i8** [[TMP70]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK-USE-PPC64LE-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK-USE-PPC64LE-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK-USE-PPC64LE-NEXT:    store i32 1, i32* [[TMP73]], align 4
+// CHECK-USE-PPC64LE-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK-USE-PPC64LE-NEXT:    store i32 1, i32* [[TMP74]], align 4
+// CHECK-USE-PPC64LE-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK-USE-PPC64LE-NEXT:    store i8** [[TMP71]], i8*** [[TMP75]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK-USE-PPC64LE-NEXT:    store i8** [[TMP72]], i8*** [[TMP76]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK-USE-PPC64LE-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.1, i32 0, i32 0), i64** [[TMP77]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK-USE-PPC64LE-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.2, i32 0, i32 0), i64** [[TMP78]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK-USE-PPC64LE-NEXT:    store i8** null, i8*** [[TMP79]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK-USE-PPC64LE-NEXT:    store i8** null, i8*** [[TMP80]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP81:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l780.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK-USE-PPC64LE-NEXT:    [[TMP82:%.*]] = icmp ne i32 [[TMP81]], 0
+// CHECK-USE-PPC64LE-NEXT:    br i1 [[TMP82]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK-USE-PPC64LE:       omp_offload.failed7:
+// CHECK-USE-PPC64LE-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l780(i32* [[A]]) #[[ATTR3]]
+// CHECK-USE-PPC64LE-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK-USE-PPC64LE:       omp_offload.cont8:
 // CHECK-USE-PPC64LE-NEXT:    ret void
 //
 // CHECK-USE-I386-LABEL: @_Z20explicit_maps_singlei(
@@ -341,30 +375,64 @@ struct ST {
 // CHECK-USE-I386-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [7 x i8*], [7 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK-USE-I386-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [7 x i8*], [7 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK-USE-I386-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [7 x i64], [7 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK-USE-I386-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l630.region_id, i32 7, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([7 x i64], [7 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null)
-// CHECK-USE-I386-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
-// CHECK-USE-I386-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK-USE-I386-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK-USE-I386-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK-USE-I386-NEXT:    store i32 1, i32* [[TMP56]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK-USE-I386-NEXT:    store i32 7, i32* [[TMP57]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK-USE-I386-NEXT:    store i8** [[TMP53]], i8*** [[TMP58]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK-USE-I386-NEXT:    store i8** [[TMP54]], i8*** [[TMP59]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK-USE-I386-NEXT:    store i64* [[TMP55]], i64** [[TMP60]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK-USE-I386-NEXT:    store i64* getelementptr inbounds ([7 x i64], [7 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP61]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK-USE-I386-NEXT:    store i8** null, i8*** [[TMP62]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK-USE-I386-NEXT:    store i8** null, i8*** [[TMP63]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP64:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l766.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK-USE-I386-NEXT:    [[TMP65:%.*]] = icmp ne i32 [[TMP64]], 0
+// CHECK-USE-I386-NEXT:    br i1 [[TMP65]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK-USE-I386:       omp_offload.failed:
-// CHECK-USE-I386-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l630(%struct.ST* [[ST1]], i32* [[A]], %struct.ST* [[ST2]]) #[[ATTR3:[0-9]+]]
+// CHECK-USE-I386-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l766(%struct.ST* [[ST1]], i32* [[A]], %struct.ST* [[ST2]]) #[[ATTR3:[0-9]+]]
 // CHECK-USE-I386-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK-USE-I386:       omp_offload.cont:
-// CHECK-USE-I386-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK-USE-I386-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to i32**
-// CHECK-USE-I386-NEXT:    store i32* [[A]], i32** [[TMP59]], align 4
-// CHECK-USE-I386-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK-USE-I386-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to i32**
-// CHECK-USE-I386-NEXT:    store i32* [[A]], i32** [[TMP61]], align 4
-// CHECK-USE-I386-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK-USE-I386-NEXT:    store i8* null, i8** [[TMP62]], align 4
-// CHECK-USE-I386-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK-USE-I386-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK-USE-I386-NEXT:    [[TMP65:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l644.region_id, i32 1, i8** [[TMP63]], i8** [[TMP64]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.1, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.2, i32 0, i32 0), i8** null, i8** null)
-// CHECK-USE-I386-NEXT:    [[TMP66:%.*]] = icmp ne i32 [[TMP65]], 0
-// CHECK-USE-I386-NEXT:    br i1 [[TMP66]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK-USE-I386:       omp_offload.failed6:
-// CHECK-USE-I386-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l644(i32* [[A]]) #[[ATTR3]]
-// CHECK-USE-I386-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK-USE-I386:       omp_offload.cont7:
+// CHECK-USE-I386-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK-USE-I386-NEXT:    [[TMP67:%.*]] = bitcast i8** [[TMP66]] to i32**
+// CHECK-USE-I386-NEXT:    store i32* [[A]], i32** [[TMP67]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK-USE-I386-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i32**
+// CHECK-USE-I386-NEXT:    store i32* [[A]], i32** [[TMP69]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK-USE-I386-NEXT:    store i8* null, i8** [[TMP70]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK-USE-I386-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK-USE-I386-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK-USE-I386-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK-USE-I386-NEXT:    store i32 1, i32* [[TMP73]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK-USE-I386-NEXT:    store i32 1, i32* [[TMP74]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK-USE-I386-NEXT:    store i8** [[TMP71]], i8*** [[TMP75]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK-USE-I386-NEXT:    store i8** [[TMP72]], i8*** [[TMP76]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK-USE-I386-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.1, i32 0, i32 0), i64** [[TMP77]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK-USE-I386-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.2, i32 0, i32 0), i64** [[TMP78]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK-USE-I386-NEXT:    store i8** null, i8*** [[TMP79]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK-USE-I386-NEXT:    store i8** null, i8*** [[TMP80]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP81:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l780.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK-USE-I386-NEXT:    [[TMP82:%.*]] = icmp ne i32 [[TMP81]], 0
+// CHECK-USE-I386-NEXT:    br i1 [[TMP82]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK-USE-I386:       omp_offload.failed7:
+// CHECK-USE-I386-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l780(i32* [[A]]) #[[ATTR3]]
+// CHECK-USE-I386-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK-USE-I386:       omp_offload.cont8:
 // CHECK-USE-I386-NEXT:    ret void
 //
 // CHECK-NOUSE-PPC64LE-LABEL: @_Z20explicit_maps_singlei(
@@ -466,30 +534,64 @@ struct ST {
 // CHECK-NOUSE-PPC64LE-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [7 x i8*], [7 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK-NOUSE-PPC64LE-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [7 x i8*], [7 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK-NOUSE-PPC64LE-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [7 x i64], [7 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l630.region_id, i32 7, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([7 x i64], [7 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null)
-// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
-// CHECK-NOUSE-PPC64LE-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK-NOUSE-PPC64LE-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK-NOUSE-PPC64LE-NEXT:    store i32 1, i32* [[TMP56]], align 4
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK-NOUSE-PPC64LE-NEXT:    store i32 7, i32* [[TMP57]], align 4
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK-NOUSE-PPC64LE-NEXT:    store i8** [[TMP53]], i8*** [[TMP58]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK-NOUSE-PPC64LE-NEXT:    store i8** [[TMP54]], i8*** [[TMP59]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK-NOUSE-PPC64LE-NEXT:    store i64* [[TMP55]], i64** [[TMP60]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK-NOUSE-PPC64LE-NEXT:    store i64* getelementptr inbounds ([7 x i64], [7 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP61]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK-NOUSE-PPC64LE-NEXT:    store i8** null, i8*** [[TMP62]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK-NOUSE-PPC64LE-NEXT:    store i8** null, i8*** [[TMP63]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP64:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l766.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP65:%.*]] = icmp ne i32 [[TMP64]], 0
+// CHECK-NOUSE-PPC64LE-NEXT:    br i1 [[TMP65]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK-NOUSE-PPC64LE:       omp_offload.failed:
-// CHECK-NOUSE-PPC64LE-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l630() #[[ATTR3:[0-9]+]]
+// CHECK-NOUSE-PPC64LE-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l766() #[[ATTR3:[0-9]+]]
 // CHECK-NOUSE-PPC64LE-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK-NOUSE-PPC64LE:       omp_offload.cont:
-// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to i32**
-// CHECK-NOUSE-PPC64LE-NEXT:    store i32* [[A]], i32** [[TMP59]], align 8
-// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to i32**
-// CHECK-NOUSE-PPC64LE-NEXT:    store i32* [[A]], i32** [[TMP61]], align 8
-// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK-NOUSE-PPC64LE-NEXT:    store i8* null, i8** [[TMP62]], align 8
-// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP65:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l644.region_id, i32 1, i8** [[TMP63]], i8** [[TMP64]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.1, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.2, i32 0, i32 0), i8** null, i8** null)
-// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP66:%.*]] = icmp ne i32 [[TMP65]], 0
-// CHECK-NOUSE-PPC64LE-NEXT:    br i1 [[TMP66]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK-NOUSE-PPC64LE:       omp_offload.failed6:
-// CHECK-NOUSE-PPC64LE-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l644() #[[ATTR3]]
-// CHECK-NOUSE-PPC64LE-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK-NOUSE-PPC64LE:       omp_offload.cont7:
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP67:%.*]] = bitcast i8** [[TMP66]] to i32**
+// CHECK-NOUSE-PPC64LE-NEXT:    store i32* [[A]], i32** [[TMP67]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i32**
+// CHECK-NOUSE-PPC64LE-NEXT:    store i32* [[A]], i32** [[TMP69]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK-NOUSE-PPC64LE-NEXT:    store i8* null, i8** [[TMP70]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK-NOUSE-PPC64LE-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK-NOUSE-PPC64LE-NEXT:    store i32 1, i32* [[TMP73]], align 4
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK-NOUSE-PPC64LE-NEXT:    store i32 1, i32* [[TMP74]], align 4
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK-NOUSE-PPC64LE-NEXT:    store i8** [[TMP71]], i8*** [[TMP75]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK-NOUSE-PPC64LE-NEXT:    store i8** [[TMP72]], i8*** [[TMP76]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK-NOUSE-PPC64LE-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.1, i32 0, i32 0), i64** [[TMP77]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK-NOUSE-PPC64LE-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.2, i32 0, i32 0), i64** [[TMP78]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK-NOUSE-PPC64LE-NEXT:    store i8** null, i8*** [[TMP79]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK-NOUSE-PPC64LE-NEXT:    store i8** null, i8*** [[TMP80]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP81:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l780.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP82:%.*]] = icmp ne i32 [[TMP81]], 0
+// CHECK-NOUSE-PPC64LE-NEXT:    br i1 [[TMP82]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK-NOUSE-PPC64LE:       omp_offload.failed7:
+// CHECK-NOUSE-PPC64LE-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l780() #[[ATTR3]]
+// CHECK-NOUSE-PPC64LE-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK-NOUSE-PPC64LE:       omp_offload.cont8:
 // CHECK-NOUSE-PPC64LE-NEXT:    ret void
 //
 // CHECK-NOUSE-I386-LABEL: @_Z20explicit_maps_singlei(
@@ -591,30 +693,64 @@ struct ST {
 // CHECK-NOUSE-I386-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [7 x i8*], [7 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK-NOUSE-I386-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [7 x i8*], [7 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK-NOUSE-I386-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [7 x i64], [7 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK-NOUSE-I386-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l630.region_id, i32 7, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([7 x i64], [7 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null)
-// CHECK-NOUSE-I386-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
-// CHECK-NOUSE-I386-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK-NOUSE-I386-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK-NOUSE-I386-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK-NOUSE-I386-NEXT:    store i32 1, i32* [[TMP56]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK-NOUSE-I386-NEXT:    store i32 7, i32* [[TMP57]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK-NOUSE-I386-NEXT:    store i8** [[TMP53]], i8*** [[TMP58]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK-NOUSE-I386-NEXT:    store i8** [[TMP54]], i8*** [[TMP59]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK-NOUSE-I386-NEXT:    store i64* [[TMP55]], i64** [[TMP60]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK-NOUSE-I386-NEXT:    store i64* getelementptr inbounds ([7 x i64], [7 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP61]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK-NOUSE-I386-NEXT:    store i8** null, i8*** [[TMP62]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK-NOUSE-I386-NEXT:    store i8** null, i8*** [[TMP63]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP64:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l766.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK-NOUSE-I386-NEXT:    [[TMP65:%.*]] = icmp ne i32 [[TMP64]], 0
+// CHECK-NOUSE-I386-NEXT:    br i1 [[TMP65]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK-NOUSE-I386:       omp_offload.failed:
-// CHECK-NOUSE-I386-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l630() #[[ATTR3:[0-9]+]]
+// CHECK-NOUSE-I386-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l766() #[[ATTR3:[0-9]+]]
 // CHECK-NOUSE-I386-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK-NOUSE-I386:       omp_offload.cont:
-// CHECK-NOUSE-I386-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK-NOUSE-I386-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to i32**
-// CHECK-NOUSE-I386-NEXT:    store i32* [[A]], i32** [[TMP59]], align 4
-// CHECK-NOUSE-I386-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK-NOUSE-I386-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to i32**
-// CHECK-NOUSE-I386-NEXT:    store i32* [[A]], i32** [[TMP61]], align 4
-// CHECK-NOUSE-I386-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK-NOUSE-I386-NEXT:    store i8* null, i8** [[TMP62]], align 4
-// CHECK-NOUSE-I386-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK-NOUSE-I386-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK-NOUSE-I386-NEXT:    [[TMP65:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l644.region_id, i32 1, i8** [[TMP63]], i8** [[TMP64]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.1, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.2, i32 0, i32 0), i8** null, i8** null)
-// CHECK-NOUSE-I386-NEXT:    [[TMP66:%.*]] = icmp ne i32 [[TMP65]], 0
-// CHECK-NOUSE-I386-NEXT:    br i1 [[TMP66]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK-NOUSE-I386:       omp_offload.failed6:
-// CHECK-NOUSE-I386-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l644() #[[ATTR3]]
-// CHECK-NOUSE-I386-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK-NOUSE-I386:       omp_offload.cont7:
+// CHECK-NOUSE-I386-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK-NOUSE-I386-NEXT:    [[TMP67:%.*]] = bitcast i8** [[TMP66]] to i32**
+// CHECK-NOUSE-I386-NEXT:    store i32* [[A]], i32** [[TMP67]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK-NOUSE-I386-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i32**
+// CHECK-NOUSE-I386-NEXT:    store i32* [[A]], i32** [[TMP69]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK-NOUSE-I386-NEXT:    store i8* null, i8** [[TMP70]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK-NOUSE-I386-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK-NOUSE-I386-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK-NOUSE-I386-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK-NOUSE-I386-NEXT:    store i32 1, i32* [[TMP73]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK-NOUSE-I386-NEXT:    store i32 1, i32* [[TMP74]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK-NOUSE-I386-NEXT:    store i8** [[TMP71]], i8*** [[TMP75]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK-NOUSE-I386-NEXT:    store i8** [[TMP72]], i8*** [[TMP76]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK-NOUSE-I386-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.1, i32 0, i32 0), i64** [[TMP77]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK-NOUSE-I386-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.2, i32 0, i32 0), i64** [[TMP78]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK-NOUSE-I386-NEXT:    store i8** null, i8*** [[TMP79]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK-NOUSE-I386-NEXT:    store i8** null, i8*** [[TMP80]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP81:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l780.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK-NOUSE-I386-NEXT:    [[TMP82:%.*]] = icmp ne i32 [[TMP81]], 0
+// CHECK-NOUSE-I386-NEXT:    br i1 [[TMP82]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK-NOUSE-I386:       omp_offload.failed7:
+// CHECK-NOUSE-I386-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z20explicit_maps_singlei_l780() #[[ATTR3]]
+// CHECK-NOUSE-I386-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK-NOUSE-I386:       omp_offload.cont8:
 // CHECK-NOUSE-I386-NEXT:    ret void
 //
 void explicit_maps_single(int ii) {
@@ -698,11 +834,28 @@ void explicit_maps_single(int ii) {
 // CHECK-USE-PPC64LE-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK-USE-PPC64LE-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK-USE-PPC64LE-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK-USE-PPC64LE-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l887.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* [[TMP26]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null)
-// CHECK-USE-PPC64LE-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK-USE-PPC64LE-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK-USE-PPC64LE-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK-USE-PPC64LE-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK-USE-PPC64LE-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK-USE-PPC64LE-NEXT:    store i32 3, i32* [[TMP28]], align 4
+// CHECK-USE-PPC64LE-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK-USE-PPC64LE-NEXT:    store i8** [[TMP24]], i8*** [[TMP29]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK-USE-PPC64LE-NEXT:    store i8** [[TMP25]], i8*** [[TMP30]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK-USE-PPC64LE-NEXT:    store i64* [[TMP26]], i64** [[TMP31]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK-USE-PPC64LE-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP32]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK-USE-PPC64LE-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK-USE-PPC64LE-NEXT:    store i8** null, i8*** [[TMP34]], align 8
+// CHECK-USE-PPC64LE-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l1091.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK-USE-PPC64LE-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK-USE-PPC64LE-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK-USE-PPC64LE:       omp_offload.failed:
-// CHECK-USE-PPC64LE-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l887(%struct.ST* [[THIS1]]) #[[ATTR3]]
+// CHECK-USE-PPC64LE-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l1091(%struct.ST* [[THIS1]]) #[[ATTR3]]
 // CHECK-USE-PPC64LE-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK-USE-PPC64LE:       omp_offload.cont:
 // CHECK-USE-PPC64LE-NEXT:    ret void
@@ -756,11 +909,28 @@ void explicit_maps_single(int ii) {
 // CHECK-USE-I386-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK-USE-I386-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK-USE-I386-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK-USE-I386-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l887.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* [[TMP26]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null)
-// CHECK-USE-I386-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK-USE-I386-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK-USE-I386-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK-USE-I386-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK-USE-I386-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK-USE-I386-NEXT:    store i32 3, i32* [[TMP28]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK-USE-I386-NEXT:    store i8** [[TMP24]], i8*** [[TMP29]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK-USE-I386-NEXT:    store i8** [[TMP25]], i8*** [[TMP30]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK-USE-I386-NEXT:    store i64* [[TMP26]], i64** [[TMP31]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK-USE-I386-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP32]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK-USE-I386-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK-USE-I386-NEXT:    store i8** null, i8*** [[TMP34]], align 4
+// CHECK-USE-I386-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l1091.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK-USE-I386-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK-USE-I386-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK-USE-I386:       omp_offload.failed:
-// CHECK-USE-I386-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l887(%struct.ST* [[THIS1]]) #[[ATTR3]]
+// CHECK-USE-I386-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l1091(%struct.ST* [[THIS1]]) #[[ATTR3]]
 // CHECK-USE-I386-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK-USE-I386:       omp_offload.cont:
 // CHECK-USE-I386-NEXT:    ret void
@@ -814,11 +984,28 @@ void explicit_maps_single(int ii) {
 // CHECK-NOUSE-PPC64LE-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK-NOUSE-PPC64LE-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK-NOUSE-PPC64LE-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l887.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* [[TMP26]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null)
-// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK-NOUSE-PPC64LE-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK-NOUSE-PPC64LE-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK-NOUSE-PPC64LE-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK-NOUSE-PPC64LE-NEXT:    store i32 3, i32* [[TMP28]], align 4
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK-NOUSE-PPC64LE-NEXT:    store i8** [[TMP24]], i8*** [[TMP29]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK-NOUSE-PPC64LE-NEXT:    store i8** [[TMP25]], i8*** [[TMP30]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK-NOUSE-PPC64LE-NEXT:    store i64* [[TMP26]], i64** [[TMP31]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK-NOUSE-PPC64LE-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP32]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK-NOUSE-PPC64LE-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK-NOUSE-PPC64LE-NEXT:    store i8** null, i8*** [[TMP34]], align 8
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l1091.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK-NOUSE-PPC64LE-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK-NOUSE-PPC64LE-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK-NOUSE-PPC64LE:       omp_offload.failed:
-// CHECK-NOUSE-PPC64LE-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l887() #[[ATTR3]]
+// CHECK-NOUSE-PPC64LE-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l1091() #[[ATTR3]]
 // CHECK-NOUSE-PPC64LE-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK-NOUSE-PPC64LE:       omp_offload.cont:
 // CHECK-NOUSE-PPC64LE-NEXT:    ret void
@@ -872,11 +1059,28 @@ void explicit_maps_single(int ii) {
 // CHECK-NOUSE-I386-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK-NOUSE-I386-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK-NOUSE-I386-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK-NOUSE-I386-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l887.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* [[TMP26]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null)
-// CHECK-NOUSE-I386-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK-NOUSE-I386-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK-NOUSE-I386-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK-NOUSE-I386-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK-NOUSE-I386-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK-NOUSE-I386-NEXT:    store i32 3, i32* [[TMP28]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK-NOUSE-I386-NEXT:    store i8** [[TMP24]], i8*** [[TMP29]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK-NOUSE-I386-NEXT:    store i8** [[TMP25]], i8*** [[TMP30]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK-NOUSE-I386-NEXT:    store i64* [[TMP26]], i64** [[TMP31]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK-NOUSE-I386-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP32]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK-NOUSE-I386-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK-NOUSE-I386-NEXT:    store i8** null, i8*** [[TMP34]], align 4
+// CHECK-NOUSE-I386-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l1091.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK-NOUSE-I386-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK-NOUSE-I386-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK-NOUSE-I386:       omp_offload.failed:
-// CHECK-NOUSE-I386-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l887() #[[ATTR3]]
+// CHECK-NOUSE-I386-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2ST20test_present_membersEv_l1091() #[[ATTR3]]
 // CHECK-NOUSE-I386-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK-NOUSE-I386:       omp_offload.cont:
 // CHECK-NOUSE-I386-NEXT:    ret void

diff --git a/clang/test/OpenMP/target_map_names.cpp b/clang/test/OpenMP/target_map_names.cpp
index ae58750..806715a 100644 (file)
--- a/clang/test/OpenMP/target_map_names.cpp
+++ b/clang/test/OpenMP/target_map_names.cpp
@@ -198,28 +198,8 @@ void secondMapNameInClause() {
   x = y = 1;
 }
 
-// DEBUG: %{{.+}} = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: %{{.+}} = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}}, i32 {{.+}}, i32 {{.+}})
-// DEBUG: call void @__tgt_target_data_begin_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: call void @__tgt_target_data_end_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: call void @__tgt_target_data_update_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: %{{.+}} = call i32 @__tgt_target_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* %{{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: %{{.+}} = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}}, i32 {{.+}}, i32 {{.+}})
-// DEBUG: call void @__tgt_target_data_begin_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: call void @__tgt_target_data_end_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: call void @__tgt_target_data_update_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-
-// CHECK: %{{.+}} = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: %{{.+}} = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}}, i32 {{.+}}, i32 {{.+}})
-// CHECK: call void @__tgt_target_data_begin_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: call void @__tgt_target_data_end_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: call void @__tgt_target_data_update_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: %{{.+}} = call i32 @__tgt_target_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* %{{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: %{{.+}} = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}}, i32 {{.+}}, i32 {{.+}})
-// CHECK: call void @__tgt_target_data_begin_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: call void @__tgt_target_data_end_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: call void @__tgt_target_data_update_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-
+// DEBUG: store i8** getelementptr inbounds ([2 x i8*], [2 x i8*]* @[[NAME:.offload_mapnames.[0-9]+]], i32 0, i32 0), i8*** %[[ARG:.+]]
+// CHECK-NOT: store i8** getelementptr inbounds ([2 x i8*], [2 x i8*]* @[[NAME:.offload_mapnames.[0-9]+]], i32 0, i32 0), i8*** %[[ARG:.+]]
 
 // DEBUG: void @.omp_mapper._ZTS2S3.id(i8* {{.*}}, i8* {{.*}}, i8* {{.*}}, i64 {{.*}}, i64 {{.*}}, i8* noundef [[NAME_ARG:%.+]])
 // DEBUG: store i8* [[NAME_ARG]], i8** [[NAME_STACK:%.+]]

diff --git a/clang/test/OpenMP/target_map_names_attr.cpp b/clang/test/OpenMP/target_map_names_attr.cpp
index 1cbba5b..82a5b56 100644 (file)
--- a/clang/test/OpenMP/target_map_names_attr.cpp
+++ b/clang/test/OpenMP/target_map_names_attr.cpp
@@ -183,28 +183,8 @@ void secondMapNameInClause() {
   x = y = 1;
 }
 
-// DEBUG: %{{.+}} = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: %{{.+}} = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}}, i32 {{.+}}, i32 {{.+}})
-// DEBUG: call void @__tgt_target_data_begin_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: call void @__tgt_target_data_end_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: call void @__tgt_target_data_update_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: %{{.+}} = call i32 @__tgt_target_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* %{{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: %{{.+}} = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}}, i32 {{.+}}, i32 {{.+}})
-// DEBUG: call void @__tgt_target_data_begin_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: call void @__tgt_target_data_end_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-// DEBUG: call void @__tgt_target_data_update_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** getelementptr inbounds ([{{[0-9]+}} x i8*], [{{[0-9]+}} x i8*]* @.offload_mapnames{{.*}}, i32 0, i32 0), i8** {{.+}})
-
-// CHECK: %{{.+}} = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: %{{.+}} = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}}, i32 {{.+}}, i32 {{.+}})
-// CHECK: call void @__tgt_target_data_begin_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: call void @__tgt_target_data_end_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: call void @__tgt_target_data_update_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: %{{.+}} = call i32 @__tgt_target_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* %{{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: %{{.+}} = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{.+}}, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}}, i32 {{.+}}, i32 {{.+}})
-// CHECK: call void @__tgt_target_data_begin_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: call void @__tgt_target_data_end_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-// CHECK: call void @__tgt_target_data_update_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* {{.+}}, i64* {{.+}}, i8** null, i8** {{.+}})
-
+// DEBUG: store i8** getelementptr inbounds ([2 x i8*], [2 x i8*]* @[[NAME:.offload_mapnames.[0-9]+]], i32 0, i32 0), i8*** %[[ARG:.+]]
+// CHECK-NOT: store i8** getelementptr inbounds ([2 x i8*], [2 x i8*]* @[[NAME:.offload_mapnames.[0-9]+]], i32 0, i32 0), i8*** %[[ARG:.+]]
 
 // DEBUG: void @.omp_mapper._ZTS2S3.id(i8* {{.*}}, i8* {{.*}}, i8* {{.*}}, i64 {{.*}}, i64 {{.*}}, i8* noundef [[NAME_ARG:%.+]])
 // DEBUG: store i8* [[NAME_ARG]], i8** [[NAME_STACK:%.+]]

diff --git a/clang/test/OpenMP/target_offload_mandatory_codegen.cpp b/clang/test/OpenMP/target_offload_mandatory_codegen.cpp
index 66fec89..fdf7a6a 100644 (file)
--- a/clang/test/OpenMP/target_offload_mandatory_codegen.cpp
+++ b/clang/test/OpenMP/target_offload_mandatory_codegen.cpp
@@ -31,9 +31,26 @@ void host_dev(int device) {
 // MANDATORY-LABEL: define {{[^@]+}}@_Z4hostv
 // MANDATORY-SAME: () #[[ATTR0]] {
 // MANDATORY-NEXT:  entry:
-// MANDATORY-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z4hostv_l12.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null)
-// MANDATORY-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// MANDATORY-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// MANDATORY-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// MANDATORY-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// MANDATORY-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// MANDATORY-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// MANDATORY-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// MANDATORY-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// MANDATORY-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// MANDATORY-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// MANDATORY-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// MANDATORY-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// MANDATORY-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// MANDATORY-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// MANDATORY-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// MANDATORY-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// MANDATORY-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// MANDATORY-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// MANDATORY-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// MANDATORY-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z4hostv_l12.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// MANDATORY-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// MANDATORY-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // MANDATORY:       omp_offload.failed:
 // MANDATORY-NEXT:    unreachable
 // MANDATORY:       omp_offload.cont:
@@ -50,9 +67,26 @@ void host_dev(int device) {
 // MANDATORY-NEXT:    [[TOBOOL:%.*]] = trunc i8 [[TMP0]] to i1
 // MANDATORY-NEXT:    br i1 [[TOBOOL]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // MANDATORY:       omp_if.then:
-// MANDATORY-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z7host_ifb_l17.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null)
-// MANDATORY-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// MANDATORY-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// MANDATORY-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// MANDATORY-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// MANDATORY-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// MANDATORY-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// MANDATORY-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// MANDATORY-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// MANDATORY-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// MANDATORY-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// MANDATORY-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// MANDATORY-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// MANDATORY-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// MANDATORY-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// MANDATORY-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// MANDATORY-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// MANDATORY-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// MANDATORY-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// MANDATORY-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// MANDATORY-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z7host_ifb_l17.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// MANDATORY-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// MANDATORY-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // MANDATORY:       omp_offload.failed:
 // MANDATORY-NEXT:    unreachable
 // MANDATORY:       omp_offload.cont:
@@ -73,9 +107,26 @@ void host_dev(int device) {
 // MANDATORY-NEXT:    store i32 [[TMP0]], i32* [[DOTCAPTURE_EXPR_]], align 4
 // MANDATORY-NEXT:    [[TMP1:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
 // MANDATORY-NEXT:    [[TMP2:%.*]] = sext i32 [[TMP1]] to i64
-// MANDATORY-NEXT:    [[TMP3:%.*]] = call i32 @__tgt_target_mapper(%struct.ident_t* @[[GLOB1]], i64 [[TMP2]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8host_devi_l22.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null)
-// MANDATORY-NEXT:    [[TMP4:%.*]] = icmp ne i32 [[TMP3]], 0
-// MANDATORY-NEXT:    br i1 [[TMP4]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// MANDATORY-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// MANDATORY-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// MANDATORY-NEXT:    store i32 1, i32* [[TMP3]], align 4
+// MANDATORY-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// MANDATORY-NEXT:    store i32 0, i32* [[TMP4]], align 4
+// MANDATORY-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// MANDATORY-NEXT:    store i8** null, i8*** [[TMP5]], align 8
+// MANDATORY-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// MANDATORY-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// MANDATORY-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// MANDATORY-NEXT:    store i64* null, i64** [[TMP7]], align 8
+// MANDATORY-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// MANDATORY-NEXT:    store i64* null, i64** [[TMP8]], align 8
+// MANDATORY-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// MANDATORY-NEXT:    store i8** null, i8*** [[TMP9]], align 8
+// MANDATORY-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// MANDATORY-NEXT:    store i8** null, i8*** [[TMP10]], align 8
+// MANDATORY-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 [[TMP2]], i32 -1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8host_devi_l22.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// MANDATORY-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
+// MANDATORY-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // MANDATORY:       omp_offload.failed:
 // MANDATORY-NEXT:    unreachable
 // MANDATORY:       omp_offload.cont:

diff --git a/clang/test/OpenMP/target_parallel_codegen.cpp b/clang/test/OpenMP/target_parallel_codegen.cpp
index 4ed22f0..9bc7093 100644 (file)
--- a/clang/test/OpenMP/target_parallel_codegen.cpp
+++ b/clang/test/OpenMP/target_parallel_codegen.cpp
@@ -314,10 +314,10 @@ int bar(int n){
 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 8
-// CHECK1-NEXT:    [[A_CASTED12:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS16:%.*]] = alloca [9 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS17:%.*]] = alloca [9 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS18:%.*]] = alloca [9 x i8*], align 8
+// CHECK1-NEXT:    [[A_CASTED13:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [9 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [9 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [9 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 8
 // CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
@@ -357,163 +357,214 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP19]], align 8
 // CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110.region_id, i32 1, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK1-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK1-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110(i64 [[TMP14]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
 // CHECK1-NEXT:    [[CONV4:%.*]] = bitcast i64* [[A_CASTED3]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP24]], i32* [[CONV4]], align 4
-// CHECK1-NEXT:    [[TMP25:%.*]] = load i64, i64* [[A_CASTED3]], align 8
-// CHECK1-NEXT:    [[TMP26:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK1-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = load i64, i64* [[A_CASTED3]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK1-NEXT:    [[CONV6:%.*]] = bitcast i64* [[AA_CASTED5]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP26]], i16* [[CONV6]], align 2
-// CHECK1-NEXT:    [[TMP27:%.*]] = load i64, i64* [[AA_CASTED5]], align 8
-// CHECK1-NEXT:    [[TMP28:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP28]], 10
+// CHECK1-NEXT:    store i16 [[TMP34]], i16* [[CONV6]], align 2
+// CHECK1-NEXT:    [[TMP35:%.*]] = load i64, i64* [[AA_CASTED5]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP36]], 10
 // CHECK1-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK1:       omp_if.then:
-// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP25]], i64* [[TMP30]], align 8
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP32:%.*]] = bitcast i8** [[TMP31]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP25]], i64* [[TMP32]], align 8
-// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP33]], align 8
-// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP27]], i64* [[TMP35]], align 8
-// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP27]], i64* [[TMP37]], align 8
-// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
-// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP41:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l119.region_id, i32 2, i8** [[TMP39]], i8** [[TMP40]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP42:%.*]] = icmp ne i32 [[TMP41]], 0
-// CHECK1-NEXT:    br i1 [[TMP42]], label [[OMP_OFFLOAD_FAILED10:%.*]], label [[OMP_OFFLOAD_CONT11:%.*]]
-// CHECK1:       omp_offload.failed10:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l119(i64 [[TMP25]], i64 [[TMP27]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT11]]
-// CHECK1:       omp_offload.cont11:
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP33]], i64* [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP33]], i64* [[TMP40]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP41]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP35]], i64* [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP35]], i64* [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS10:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP49]], align 4
+// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP50]], align 4
+// CHECK1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 8
+// CHECK1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP48]], i8*** [[TMP52]], align 8
+// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK1-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP54]], align 8
+// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP56]], align 8
+// CHECK1-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l119.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]])
+// CHECK1-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
+// CHECK1-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
+// CHECK1:       omp_offload.failed11:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l119(i64 [[TMP33]], i64 [[TMP35]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
+// CHECK1:       omp_offload.cont12:
 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK1:       omp_if.else:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l119(i64 [[TMP25]], i64 [[TMP27]]) #[[ATTR4]]
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l119(i64 [[TMP33]], i64 [[TMP35]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP43:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    [[CONV13:%.*]] = bitcast i64* [[A_CASTED12]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP43]], i32* [[CONV13]], align 4
-// CHECK1-NEXT:    [[TMP44:%.*]] = load i64, i64* [[A_CASTED12]], align 8
-// CHECK1-NEXT:    [[TMP45:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP14:%.*]] = icmp sgt i32 [[TMP45]], 20
-// CHECK1-NEXT:    br i1 [[CMP14]], label [[OMP_IF_THEN15:%.*]], label [[OMP_IF_ELSE21:%.*]]
-// CHECK1:       omp_if.then15:
-// CHECK1-NEXT:    [[TMP46:%.*]] = mul nuw i64 [[TMP2]], 4
-// CHECK1-NEXT:    [[TMP47:%.*]] = mul nuw i64 5, [[TMP5]]
-// CHECK1-NEXT:    [[TMP48:%.*]] = mul nuw i64 [[TMP47]], 8
-// CHECK1-NEXT:    [[TMP49:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP49]], i8* align 8 bitcast ([9 x i64]* @.offload_sizes.8 to i8*), i64 72, i1 false)
-// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP44]], i64* [[TMP51]], align 8
-// CHECK1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP44]], i64* [[TMP53]], align 8
-// CHECK1-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP54]], align 8
-// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to [10 x float]**
-// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP56]], align 8
-// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to [10 x float]**
-// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP58]], align 8
-// CHECK1-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP59]], align 8
-// CHECK1-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP61]], align 8
-// CHECK1-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP63:%.*]] = bitcast i8** [[TMP62]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP63]], align 8
-// CHECK1-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP64]], align 8
-// CHECK1-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP66:%.*]] = bitcast i8** [[TMP65]] to float**
-// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP66]], align 8
-// CHECK1-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to float**
-// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP68]], align 8
-// CHECK1-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK1-NEXT:    store i64 [[TMP46]], i64* [[TMP69]], align 8
-// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 3
+// CHECK1-NEXT:    [[TMP59:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[CONV14:%.*]] = bitcast i64* [[A_CASTED13]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP59]], i32* [[CONV14]], align 4
+// CHECK1-NEXT:    [[TMP60:%.*]] = load i64, i64* [[A_CASTED13]], align 8
+// CHECK1-NEXT:    [[TMP61:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP15:%.*]] = icmp sgt i32 [[TMP61]], 20
+// CHECK1-NEXT:    br i1 [[CMP15]], label [[OMP_IF_THEN16:%.*]], label [[OMP_IF_ELSE23:%.*]]
+// CHECK1:       omp_if.then16:
+// CHECK1-NEXT:    [[TMP62:%.*]] = mul nuw i64 [[TMP2]], 4
+// CHECK1-NEXT:    [[TMP63:%.*]] = mul nuw i64 5, [[TMP5]]
+// CHECK1-NEXT:    [[TMP64:%.*]] = mul nuw i64 [[TMP63]], 8
+// CHECK1-NEXT:    [[TMP65:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP65]], i8* align 8 bitcast ([9 x i64]* @.offload_sizes.8 to i8*), i64 72, i1 false)
+// CHECK1-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP67:%.*]] = bitcast i8** [[TMP66]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP60]], i64* [[TMP67]], align 8
+// CHECK1-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP60]], i64* [[TMP69]], align 8
+// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 0
 // CHECK1-NEXT:    store i8* null, i8** [[TMP70]], align 8
-// CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to [5 x [10 x double]]**
-// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP72]], align 8
-// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to [5 x [10 x double]]**
-// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP74]], align 8
-// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 4
+// CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to [10 x float]**
+// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP72]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to [10 x float]**
+// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP74]], align 8
+// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 1
 // CHECK1-NEXT:    store i8* null, i8** [[TMP75]], align 8
-// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to i64*
-// CHECK1-NEXT:    store i64 5, i64* [[TMP77]], align 8
-// CHECK1-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 5
+// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP77]], align 8
+// CHECK1-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i64*
-// CHECK1-NEXT:    store i64 5, i64* [[TMP79]], align 8
-// CHECK1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 5
+// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP79]], align 8
+// CHECK1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 2
 // CHECK1-NEXT:    store i8* null, i8** [[TMP80]], align 8
-// CHECK1-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP82]], align 8
-// CHECK1-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP84]], align 8
-// CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 6
-// CHECK1-NEXT:    store i8* null, i8** [[TMP85]], align 8
-// CHECK1-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to double**
-// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP87]], align 8
-// CHECK1-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to double**
-// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP89]], align 8
-// CHECK1-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK1-NEXT:    store i64 [[TMP48]], i64* [[TMP90]], align 8
-// CHECK1-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 7
+// CHECK1-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to float**
+// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP82]], align 8
+// CHECK1-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to float**
+// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP84]], align 8
+// CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK1-NEXT:    store i64 [[TMP62]], i64* [[TMP85]], align 8
+// CHECK1-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 3
+// CHECK1-NEXT:    store i8* null, i8** [[TMP86]], align 8
+// CHECK1-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP88:%.*]] = bitcast i8** [[TMP87]] to [5 x [10 x double]]**
+// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP88]], align 8
+// CHECK1-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to [5 x [10 x double]]**
+// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP90]], align 8
+// CHECK1-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 4
 // CHECK1-NEXT:    store i8* null, i8** [[TMP91]], align 8
-// CHECK1-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to %struct.TT**
-// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP93]], align 8
-// CHECK1-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to %struct.TT**
-// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP95]], align 8
-// CHECK1-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 8
+// CHECK1-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to i64*
+// CHECK1-NEXT:    store i64 5, i64* [[TMP93]], align 8
+// CHECK1-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i64*
+// CHECK1-NEXT:    store i64 5, i64* [[TMP95]], align 8
+// CHECK1-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 5
 // CHECK1-NEXT:    store i8* null, i8** [[TMP96]], align 8
-// CHECK1-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP100:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l144.region_id, i32 9, i8** [[TMP97]], i8** [[TMP98]], i64* [[TMP99]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP101:%.*]] = icmp ne i32 [[TMP100]], 0
-// CHECK1-NEXT:    br i1 [[TMP101]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
-// CHECK1:       omp_offload.failed19:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l144(i64 [[TMP44]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
-// CHECK1:       omp_offload.cont20:
-// CHECK1-NEXT:    br label [[OMP_IF_END22:%.*]]
-// CHECK1:       omp_if.else21:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l144(i64 [[TMP44]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_IF_END22]]
-// CHECK1:       omp_if.end22:
-// CHECK1-NEXT:    [[TMP102:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    [[TMP103:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP103]])
-// CHECK1-NEXT:    ret i32 [[TMP102]]
+// CHECK1-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 6
+// CHECK1-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP98]], align 8
+// CHECK1-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 6
+// CHECK1-NEXT:    [[TMP100:%.*]] = bitcast i8** [[TMP99]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP100]], align 8
+// CHECK1-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 6
+// CHECK1-NEXT:    store i8* null, i8** [[TMP101]], align 8
+// CHECK1-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to double**
+// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP103]], align 8
+// CHECK1-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to double**
+// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP105]], align 8
+// CHECK1-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK1-NEXT:    store i64 [[TMP64]], i64* [[TMP106]], align 8
+// CHECK1-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 7
+// CHECK1-NEXT:    store i8* null, i8** [[TMP107]], align 8
+// CHECK1-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to %struct.TT**
+// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP109]], align 8
+// CHECK1-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP111:%.*]] = bitcast i8** [[TMP110]] to %struct.TT**
+// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP111]], align 8
+// CHECK1-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 8
+// CHECK1-NEXT:    store i8* null, i8** [[TMP112]], align 8
+// CHECK1-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS20:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP116]], align 4
+// CHECK1-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 9, i32* [[TMP117]], align 4
+// CHECK1-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP113]], i8*** [[TMP118]], align 8
+// CHECK1-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP114]], i8*** [[TMP119]], align 8
+// CHECK1-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP115]], i64** [[TMP120]], align 8
+// CHECK1-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP121]], align 8
+// CHECK1-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP122]], align 8
+// CHECK1-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP123]], align 8
+// CHECK1-NEXT:    [[TMP124:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l144.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]])
+// CHECK1-NEXT:    [[TMP125:%.*]] = icmp ne i32 [[TMP124]], 0
+// CHECK1-NEXT:    br i1 [[TMP125]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
+// CHECK1:       omp_offload.failed21:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l144(i64 [[TMP60]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
+// CHECK1:       omp_offload.cont22:
+// CHECK1-NEXT:    br label [[OMP_IF_END24:%.*]]
+// CHECK1:       omp_if.else23:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l144(i64 [[TMP60]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_IF_END24]]
+// CHECK1:       omp_if.end24:
+// CHECK1-NEXT:    [[TMP126:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP127:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP127]])
+// CHECK1-NEXT:    ret i32 [[TMP126]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l100
@@ -542,6 +593,7 @@ int bar(int n){
 // CHECK1-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
 // CHECK1-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
 // CHECK1-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 8
+// CHECK1-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK1-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -565,9 +617,25 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !21
 // CHECK1-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !21
 // CHECK1-NEXT:    [[TMP10:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !21
-// CHECK1-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l100.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK1-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK1-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP11]], align 4, !noalias !21
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP12]], align 4, !noalias !21
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP16]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP18]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP19:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l100.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK1-NEXT:    [[TMP20:%.*]] = icmp ne i32 [[TMP19]], 0
+// CHECK1-NEXT:    br i1 [[TMP20]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK1:       omp_offload.failed.i:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l100() #[[ATTR4]]
 // CHECK1-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
@@ -923,9 +991,26 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK1-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK1-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -935,15 +1020,15 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP42:%.*]] = mul nsw i64 1, [[TMP2]]
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP42]]
+// CHECK1-NEXT:    [[TMP50:%.*]] = mul nsw i64 1, [[TMP2]]
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP50]]
 // CHECK1-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
-// CHECK1-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK1-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP44]]
-// CHECK1-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK1-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP52]]
+// CHECK1-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK1-NEXT:    ret i32 [[ADD4]]
 //
 //
@@ -1015,9 +1100,26 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
 // CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l198.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK1-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l198.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK1-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l198(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1027,8 +1129,8 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l198(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP31]]
+// CHECK1-NEXT:    [[TMP39:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP39]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -1084,9 +1186,26 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP19]], align 8
 // CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l181.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK1-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l181.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK1-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l181(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1096,8 +1215,8 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l181(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP24]]
+// CHECK1-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216
@@ -1326,10 +1445,10 @@ int bar(int n){
 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS5:%.*]] = alloca [2 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS6:%.*]] = alloca [2 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS7:%.*]] = alloca [2 x i8*], align 4
-// CHECK3-NEXT:    [[A_CASTED10:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [9 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [9 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [9 x i8*], align 4
+// CHECK3-NEXT:    [[A_CASTED11:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS14:%.*]] = alloca [9 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS15:%.*]] = alloca [9 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS16:%.*]] = alloca [9 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 4
 // CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
@@ -1366,163 +1485,214 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP17]], align 4
 // CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110.region_id, i32 1, i8** [[TMP18]], i8** [[TMP19]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
-// CHECK3-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP18]], i8*** [[TMP22]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP19]], i8*** [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
+// CHECK3-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110(i32 [[TMP12]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP22:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP22]], i32* [[A_CASTED2]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = load i32, i32* [[A_CASTED2]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP30]], i32* [[A_CASTED2]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A_CASTED2]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK3-NEXT:    [[CONV4:%.*]] = bitcast i32* [[AA_CASTED3]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP24]], i16* [[CONV4]], align 2
-// CHECK3-NEXT:    [[TMP25:%.*]] = load i32, i32* [[AA_CASTED3]], align 4
-// CHECK3-NEXT:    [[TMP26:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP26]], 10
+// CHECK3-NEXT:    store i16 [[TMP32]], i16* [[CONV4]], align 2
+// CHECK3-NEXT:    [[TMP33:%.*]] = load i32, i32* [[AA_CASTED3]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP34]], 10
 // CHECK3-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK3:       omp_if.then:
-// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP23]], i32* [[TMP28]], align 4
-// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP23]], i32* [[TMP30]], align 4
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP31]], align 4
-// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP33:%.*]] = bitcast i8** [[TMP32]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP25]], i32* [[TMP33]], align 4
-// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP25]], i32* [[TMP35]], align 4
-// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP36]], align 4
-// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l119.region_id, i32 2, i8** [[TMP37]], i8** [[TMP38]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
-// CHECK3-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
-// CHECK3:       omp_offload.failed8:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l119(i32 [[TMP23]], i32 [[TMP25]]) #[[ATTR4]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
-// CHECK3:       omp_offload.cont9:
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP31]], i32* [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP31]], i32* [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP39]], align 4
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP33]], i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP33]], i32* [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS8:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP45]], i8*** [[TMP49]], align 4
+// CHECK3-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP46]], i8*** [[TMP50]], align 4
+// CHECK3-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP51]], align 4
+// CHECK3-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP52]], align 4
+// CHECK3-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP53]], align 4
+// CHECK3-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP54]], align 4
+// CHECK3-NEXT:    [[TMP55:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l119.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]])
+// CHECK3-NEXT:    [[TMP56:%.*]] = icmp ne i32 [[TMP55]], 0
+// CHECK3-NEXT:    br i1 [[TMP56]], label [[OMP_OFFLOAD_FAILED9:%.*]], label [[OMP_OFFLOAD_CONT10:%.*]]
+// CHECK3:       omp_offload.failed9:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l119(i32 [[TMP31]], i32 [[TMP33]]) #[[ATTR4]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT10]]
+// CHECK3:       omp_offload.cont10:
 // CHECK3-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK3:       omp_if.else:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l119(i32 [[TMP23]], i32 [[TMP25]]) #[[ATTR4]]
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l119(i32 [[TMP31]], i32 [[TMP33]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP41:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP41]], i32* [[A_CASTED10]], align 4
-// CHECK3-NEXT:    [[TMP42:%.*]] = load i32, i32* [[A_CASTED10]], align 4
-// CHECK3-NEXT:    [[TMP43:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP11:%.*]] = icmp sgt i32 [[TMP43]], 20
-// CHECK3-NEXT:    br i1 [[CMP11]], label [[OMP_IF_THEN12:%.*]], label [[OMP_IF_ELSE18:%.*]]
-// CHECK3:       omp_if.then12:
-// CHECK3-NEXT:    [[TMP44:%.*]] = mul nuw i32 [[TMP1]], 4
-// CHECK3-NEXT:    [[TMP45:%.*]] = sext i32 [[TMP44]] to i64
-// CHECK3-NEXT:    [[TMP46:%.*]] = mul nuw i32 5, [[TMP3]]
-// CHECK3-NEXT:    [[TMP47:%.*]] = mul nuw i32 [[TMP46]], 8
-// CHECK3-NEXT:    [[TMP48:%.*]] = sext i32 [[TMP47]] to i64
-// CHECK3-NEXT:    [[TMP49:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP49]], i8* align 4 bitcast ([9 x i64]* @.offload_sizes.8 to i8*), i32 72, i1 false)
-// CHECK3-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP42]], i32* [[TMP51]], align 4
-// CHECK3-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP42]], i32* [[TMP53]], align 4
-// CHECK3-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP54]], align 4
-// CHECK3-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to [10 x float]**
-// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP56]], align 4
-// CHECK3-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to [10 x float]**
-// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP58]], align 4
-// CHECK3-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP59]], align 4
-// CHECK3-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP61]], align 4
-// CHECK3-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP63:%.*]] = bitcast i8** [[TMP62]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP63]], align 4
-// CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP64]], align 4
-// CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP66:%.*]] = bitcast i8** [[TMP65]] to float**
-// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP66]], align 4
-// CHECK3-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to float**
-// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP68]], align 4
-// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK3-NEXT:    store i64 [[TMP45]], i64* [[TMP69]], align 4
-// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP57:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP57]], i32* [[A_CASTED11]], align 4
+// CHECK3-NEXT:    [[TMP58:%.*]] = load i32, i32* [[A_CASTED11]], align 4
+// CHECK3-NEXT:    [[TMP59:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP12:%.*]] = icmp sgt i32 [[TMP59]], 20
+// CHECK3-NEXT:    br i1 [[CMP12]], label [[OMP_IF_THEN13:%.*]], label [[OMP_IF_ELSE20:%.*]]
+// CHECK3:       omp_if.then13:
+// CHECK3-NEXT:    [[TMP60:%.*]] = mul nuw i32 [[TMP1]], 4
+// CHECK3-NEXT:    [[TMP61:%.*]] = sext i32 [[TMP60]] to i64
+// CHECK3-NEXT:    [[TMP62:%.*]] = mul nuw i32 5, [[TMP3]]
+// CHECK3-NEXT:    [[TMP63:%.*]] = mul nuw i32 [[TMP62]], 8
+// CHECK3-NEXT:    [[TMP64:%.*]] = sext i32 [[TMP63]] to i64
+// CHECK3-NEXT:    [[TMP65:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP65]], i8* align 4 bitcast ([9 x i64]* @.offload_sizes.8 to i8*), i32 72, i1 false)
+// CHECK3-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP67:%.*]] = bitcast i8** [[TMP66]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP58]], i32* [[TMP67]], align 4
+// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP58]], i32* [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 0
 // CHECK3-NEXT:    store i8* null, i8** [[TMP70]], align 4
-// CHECK3-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to [5 x [10 x double]]**
-// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP72]], align 4
-// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to [5 x [10 x double]]**
-// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP74]], align 4
-// CHECK3-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to [10 x float]**
+// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP72]], align 4
+// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to [10 x float]**
+// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP74]], align 4
+// CHECK3-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 1
 // CHECK3-NEXT:    store i8* null, i8** [[TMP75]], align 4
-// CHECK3-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to i32*
-// CHECK3-NEXT:    store i32 5, i32* [[TMP77]], align 4
-// CHECK3-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 5
+// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP77]], align 4
+// CHECK3-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
-// CHECK3-NEXT:    store i32 5, i32* [[TMP79]], align 4
-// CHECK3-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP79]], align 4
+// CHECK3-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 2
 // CHECK3-NEXT:    store i8* null, i8** [[TMP80]], align 4
-// CHECK3-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 6
-// CHECK3-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP82]], align 4
-// CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 6
-// CHECK3-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP84]], align 4
-// CHECK3-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 6
-// CHECK3-NEXT:    store i8* null, i8** [[TMP85]], align 4
-// CHECK3-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 7
-// CHECK3-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to double**
-// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP87]], align 4
-// CHECK3-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 7
-// CHECK3-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to double**
-// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP89]], align 4
-// CHECK3-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK3-NEXT:    store i64 [[TMP48]], i64* [[TMP90]], align 4
-// CHECK3-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to float**
+// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP82]], align 4
+// CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to float**
+// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP84]], align 4
+// CHECK3-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK3-NEXT:    store i64 [[TMP61]], i64* [[TMP85]], align 4
+// CHECK3-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 3
+// CHECK3-NEXT:    store i8* null, i8** [[TMP86]], align 4
+// CHECK3-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP88:%.*]] = bitcast i8** [[TMP87]] to [5 x [10 x double]]**
+// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP88]], align 4
+// CHECK3-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to [5 x [10 x double]]**
+// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP90]], align 4
+// CHECK3-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 4
 // CHECK3-NEXT:    store i8* null, i8** [[TMP91]], align 4
-// CHECK3-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 8
-// CHECK3-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to %struct.TT**
-// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP93]], align 4
-// CHECK3-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 8
-// CHECK3-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to %struct.TT**
-// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP95]], align 4
-// CHECK3-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to i32*
+// CHECK3-NEXT:    store i32 5, i32* [[TMP93]], align 4
+// CHECK3-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32*
+// CHECK3-NEXT:    store i32 5, i32* [[TMP95]], align 4
+// CHECK3-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 5
 // CHECK3-NEXT:    store i8* null, i8** [[TMP96]], align 4
-// CHECK3-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP100:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l144.region_id, i32 9, i8** [[TMP97]], i8** [[TMP98]], i64* [[TMP99]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP101:%.*]] = icmp ne i32 [[TMP100]], 0
-// CHECK3-NEXT:    br i1 [[TMP101]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK3:       omp_offload.failed16:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l144(i32 [[TMP42]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK3:       omp_offload.cont17:
-// CHECK3-NEXT:    br label [[OMP_IF_END19:%.*]]
-// CHECK3:       omp_if.else18:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l144(i32 [[TMP42]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
-// CHECK3-NEXT:    br label [[OMP_IF_END19]]
-// CHECK3:       omp_if.end19:
-// CHECK3-NEXT:    [[TMP102:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    [[TMP103:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP103]])
-// CHECK3-NEXT:    ret i32 [[TMP102]]
+// CHECK3-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 6
+// CHECK3-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP98]], align 4
+// CHECK3-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 6
+// CHECK3-NEXT:    [[TMP100:%.*]] = bitcast i8** [[TMP99]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP100]], align 4
+// CHECK3-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 6
+// CHECK3-NEXT:    store i8* null, i8** [[TMP101]], align 4
+// CHECK3-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to double**
+// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP103]], align 4
+// CHECK3-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to double**
+// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP105]], align 4
+// CHECK3-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK3-NEXT:    store i64 [[TMP64]], i64* [[TMP106]], align 4
+// CHECK3-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 7
+// CHECK3-NEXT:    store i8* null, i8** [[TMP107]], align 4
+// CHECK3-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to %struct.TT**
+// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP109]], align 4
+// CHECK3-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP111:%.*]] = bitcast i8** [[TMP110]] to %struct.TT**
+// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP111]], align 4
+// CHECK3-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 8
+// CHECK3-NEXT:    store i8* null, i8** [[TMP112]], align 4
+// CHECK3-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS17:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP116]], align 4
+// CHECK3-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 9, i32* [[TMP117]], align 4
+// CHECK3-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP113]], i8*** [[TMP118]], align 4
+// CHECK3-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP114]], i8*** [[TMP119]], align 4
+// CHECK3-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP115]], i64** [[TMP120]], align 4
+// CHECK3-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP121]], align 4
+// CHECK3-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP122]], align 4
+// CHECK3-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP123]], align 4
+// CHECK3-NEXT:    [[TMP124:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l144.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]])
+// CHECK3-NEXT:    [[TMP125:%.*]] = icmp ne i32 [[TMP124]], 0
+// CHECK3-NEXT:    br i1 [[TMP125]], label [[OMP_OFFLOAD_FAILED18:%.*]], label [[OMP_OFFLOAD_CONT19:%.*]]
+// CHECK3:       omp_offload.failed18:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l144(i32 [[TMP58]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT19]]
+// CHECK3:       omp_offload.cont19:
+// CHECK3-NEXT:    br label [[OMP_IF_END21:%.*]]
+// CHECK3:       omp_if.else20:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l144(i32 [[TMP58]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
+// CHECK3-NEXT:    br label [[OMP_IF_END21]]
+// CHECK3:       omp_if.end21:
+// CHECK3-NEXT:    [[TMP126:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    [[TMP127:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP127]])
+// CHECK3-NEXT:    ret i32 [[TMP126]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l100
@@ -1551,6 +1721,7 @@ int bar(int n){
 // CHECK3-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 4
 // CHECK3-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 4
 // CHECK3-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 4
+// CHECK3-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK3-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 4
 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -1574,9 +1745,25 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 4, !noalias !22
 // CHECK3-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !22
 // CHECK3-NEXT:    [[TMP10:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !22
-// CHECK3-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l100.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK3-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK3-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP11]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP12]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP15]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP16]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP18]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP19:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l100.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK3-NEXT:    [[TMP20:%.*]] = icmp ne i32 [[TMP19]], 0
+// CHECK3-NEXT:    br i1 [[TMP20]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK3:       omp_offload.failed.i:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l100() #[[ATTR4]]
 // CHECK3-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
@@ -1922,9 +2109,26 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK3-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK3-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1934,15 +2138,15 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP42:%.*]] = mul nsw i32 1, [[TMP1]]
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP42]]
+// CHECK3-NEXT:    [[TMP50:%.*]] = mul nsw i32 1, [[TMP1]]
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP50]]
 // CHECK3-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
-// CHECK3-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK3-NEXT:    [[CONV:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK3-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP44]]
-// CHECK3-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK3-NEXT:    [[CONV:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK3-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP52]]
+// CHECK3-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK3-NEXT:    ret i32 [[ADD3]]
 //
 //
@@ -2013,9 +2217,26 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP26]], align 4
 // CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l198.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK3-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l198.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK3-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l198(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2025,8 +2246,8 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l198(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP31]]
+// CHECK3-NEXT:    [[TMP39:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP39]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -2081,9 +2302,26 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP19]], align 4
 // CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l181.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK3-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l181.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK3-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l181(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2093,8 +2331,8 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l181(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP24]]
+// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216

diff --git a/clang/test/OpenMP/target_parallel_depend_codegen.cpp b/clang/test/OpenMP/target_parallel_depend_codegen.cpp
index 97832c7..9a4b892 100644 (file)
--- a/clang/test/OpenMP/target_parallel_depend_codegen.cpp
+++ b/clang/test/OpenMP/target_parallel_depend_codegen.cpp
@@ -187,7 +187,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 0
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // CHECK:       [[FAIL]]
@@ -205,7 +205,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 2
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 2, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SZT]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 0, i32 0, i8* null, i32 0, i8* null)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 
 // CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]

diff --git a/clang/test/OpenMP/target_parallel_for_codegen.cpp b/clang/test/OpenMP/target_parallel_for_codegen.cpp
index 80bd060..fc69858 100644 (file)
--- a/clang/test/OpenMP/target_parallel_for_codegen.cpp
+++ b/clang/test/OpenMP/target_parallel_for_codegen.cpp
@@ -351,11 +351,11 @@ int bar(int n){
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [2 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [2 x i8*], align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[A_CASTED15:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[A_CASTED16:%.*]] = alloca i64, align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [10 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [10 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [10 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [10 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [10 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [10 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [10 x i64], align 8
 // CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
@@ -372,259 +372,310 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP6:%.*]] = mul nuw i64 5, [[TMP5]]
 // CHECK1-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP6]], align 8
 // CHECK1-NEXT:    store i64 [[TMP5]], i64* [[__VLA_EXPR1]], align 8
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103() #[[ATTR4:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[CALL:%.*]] = call noundef i64 @_Z7get_valv()
 // CHECK1-NEXT:    store i64 [[CALL]], i64* [[K]], align 8
-// CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* [[A]], align 4
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP9]], i32* [[CONV]], align 4
-// CHECK1-NEXT:    [[TMP10:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = load i64, i64* [[K]], align 8
-// CHECK1-NEXT:    store i64 [[TMP11]], i64* [[K_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP12:%.*]] = load i64, i64* [[K_CASTED]], align 8
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110(i64 [[TMP10]], i64 [[TMP12]]) #[[ATTR4]]
+// CHECK1-NEXT:    store i32 [[TMP17]], i32* [[CONV]], align 4
+// CHECK1-NEXT:    [[TMP18:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = load i64, i64* [[K]], align 8
+// CHECK1-NEXT:    store i64 [[TMP19]], i64* [[K_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = load i64, i64* [[K_CASTED]], align 8
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110(i64 [[TMP18]], i64 [[TMP20]]) #[[ATTR4]]
 // CHECK1-NEXT:    store i32 12, i32* [[LIN]], align 4
-// CHECK1-NEXT:    [[TMP13:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK1-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP13]], i16* [[CONV2]], align 2
-// CHECK1-NEXT:    [[TMP14:%.*]] = load i64, i64* [[AA_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP15:%.*]] = load i32, i32* [[LIN]], align 4
+// CHECK1-NEXT:    store i16 [[TMP21]], i16* [[CONV2]], align 2
+// CHECK1-NEXT:    [[TMP22:%.*]] = load i64, i64* [[AA_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP23:%.*]] = load i32, i32* [[LIN]], align 4
 // CHECK1-NEXT:    [[CONV3:%.*]] = bitcast i64* [[LIN_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP15]], i32* [[CONV3]], align 4
-// CHECK1-NEXT:    [[TMP16:%.*]] = load i64, i64* [[LIN_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    store i32 [[TMP23]], i32* [[CONV3]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = load i64, i64* [[LIN_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = load i32, i32* [[A]], align 4
 // CHECK1-NEXT:    [[CONV5:%.*]] = bitcast i64* [[A_CASTED4]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP17]], i32* [[CONV5]], align 4
-// CHECK1-NEXT:    [[TMP18:%.*]] = load i64, i64* [[A_CASTED4]], align 8
-// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP20]], align 8
-// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP23]], align 8
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP16]], i64* [[TMP25]], align 8
-// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP16]], i64* [[TMP27]], align 8
-// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP28]], align 8
-// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK1-NEXT:    store i32 [[TMP25]], i32* [[CONV5]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = load i64, i64* [[A_CASTED4]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP22]], i64* [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP18]], i64* [[TMP30]], align 8
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP32:%.*]] = bitcast i8** [[TMP31]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP18]], i64* [[TMP32]], align 8
-// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP33]], align 8
-// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP37:%.*]] = load i16, i16* [[AA]], align 2
-// CHECK1-NEXT:    store i16 [[TMP37]], i16* [[TMP36]], align 4
-// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP39:%.*]] = load i32, i32* [[LIN]], align 4
-// CHECK1-NEXT:    store i32 [[TMP39]], i32* [[TMP38]], align 4
-// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP41:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    store i32 [[TMP41]], i32* [[TMP40]], align 4
-// CHECK1-NEXT:    [[TMP42:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i32 1, i64 120, i64 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
-// CHECK1-NEXT:    [[TMP43:%.*]] = bitcast i8* [[TMP42]] to %struct.kmp_task_t_with_privates*
-// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP43]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP44]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP46:%.*]] = load i8*, i8** [[TMP45]], align 8
-// CHECK1-NEXT:    [[TMP47:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
-// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP46]], i8* align 4 [[TMP47]], i64 12, i1 false)
-// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP43]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP49:%.*]] = bitcast i8* [[TMP46]] to %struct.anon*
-// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP48]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP51:%.*]] = bitcast [3 x i8*]* [[TMP50]] to i8*
-// CHECK1-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP34]] to i8*
-// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP51]], i8* align 8 [[TMP52]], i64 24, i1 false)
-// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP48]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP54:%.*]] = bitcast [3 x i8*]* [[TMP53]] to i8*
-// CHECK1-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP35]] to i8*
-// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP54]], i8* align 8 [[TMP55]], i64 24, i1 false)
-// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP48]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP57:%.*]] = bitcast [3 x i64]* [[TMP56]] to i8*
-// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP57]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes to i8*), i64 24, i1 false)
-// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP48]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP59:%.*]] = load i16, i16* [[AA]], align 2
-// CHECK1-NEXT:    store i16 [[TMP59]], i16* [[TMP58]], align 8
-// CHECK1-NEXT:    [[TMP60:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i8* [[TMP42]])
-// CHECK1-NEXT:    [[TMP61:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    store i64 [[TMP22]], i64* [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP33:%.*]] = bitcast i8** [[TMP32]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP24]], i64* [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP24]], i64* [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP26]], i64* [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP26]], i64* [[TMP40]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
+// CHECK1-NEXT:    store i8* null, i8** [[TMP41]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP45:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK1-NEXT:    store i16 [[TMP45]], i16* [[TMP44]], align 4
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP47:%.*]] = load i32, i32* [[LIN]], align 4
+// CHECK1-NEXT:    store i32 [[TMP47]], i32* [[TMP46]], align 4
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP49:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    store i32 [[TMP49]], i32* [[TMP48]], align 4
+// CHECK1-NEXT:    [[TMP50:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i32 1, i64 120, i64 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
+// CHECK1-NEXT:    [[TMP51:%.*]] = bitcast i8* [[TMP50]] to %struct.kmp_task_t_with_privates*
+// CHECK1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP51]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP52]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP54:%.*]] = load i8*, i8** [[TMP53]], align 8
+// CHECK1-NEXT:    [[TMP55:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
+// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP54]], i8* align 4 [[TMP55]], i64 12, i1 false)
+// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP51]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP57:%.*]] = bitcast i8* [[TMP54]] to %struct.anon*
+// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP56]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP59:%.*]] = bitcast [3 x i8*]* [[TMP58]] to i8*
+// CHECK1-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP42]] to i8*
+// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP59]], i8* align 8 [[TMP60]], i64 24, i1 false)
+// CHECK1-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP56]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP62:%.*]] = bitcast [3 x i8*]* [[TMP61]] to i8*
+// CHECK1-NEXT:    [[TMP63:%.*]] = bitcast i8** [[TMP43]] to i8*
+// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP62]], i8* align 8 [[TMP63]], i64 24, i1 false)
+// CHECK1-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP56]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP65:%.*]] = bitcast [3 x i64]* [[TMP64]] to i8*
+// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP65]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes to i8*), i64 24, i1 false)
+// CHECK1-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP56]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP67:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK1-NEXT:    store i16 [[TMP67]], i16* [[TMP66]], align 8
+// CHECK1-NEXT:    [[TMP68:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i8* [[TMP50]])
+// CHECK1-NEXT:    [[TMP69:%.*]] = load i32, i32* [[A]], align 4
 // CHECK1-NEXT:    [[CONV7:%.*]] = bitcast i64* [[A_CASTED6]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP61]], i32* [[CONV7]], align 4
-// CHECK1-NEXT:    [[TMP62:%.*]] = load i64, i64* [[A_CASTED6]], align 8
-// CHECK1-NEXT:    [[TMP63:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK1-NEXT:    store i32 [[TMP69]], i32* [[CONV7]], align 4
+// CHECK1-NEXT:    [[TMP70:%.*]] = load i64, i64* [[A_CASTED6]], align 8
+// CHECK1-NEXT:    [[TMP71:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK1-NEXT:    [[CONV9:%.*]] = bitcast i64* [[AA_CASTED8]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP63]], i16* [[CONV9]], align 2
-// CHECK1-NEXT:    [[TMP64:%.*]] = load i64, i64* [[AA_CASTED8]], align 8
-// CHECK1-NEXT:    [[TMP65:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP65]], 10
+// CHECK1-NEXT:    store i16 [[TMP71]], i16* [[CONV9]], align 2
+// CHECK1-NEXT:    [[TMP72:%.*]] = load i64, i64* [[AA_CASTED8]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP73]], 10
 // CHECK1-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK1:       omp_if.then:
-// CHECK1-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP67:%.*]] = bitcast i8** [[TMP66]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP62]], i64* [[TMP67]], align 8
-// CHECK1-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP62]], i64* [[TMP69]], align 8
-// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP70]], align 8
-// CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP64]], i64* [[TMP72]], align 8
-// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP64]], i64* [[TMP74]], align 8
-// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP75]], align 8
-// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP78:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146.region_id, i32 2, i8** [[TMP76]], i8** [[TMP77]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP79:%.*]] = icmp ne i32 [[TMP78]], 0
-// CHECK1-NEXT:    br i1 [[TMP79]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
-// CHECK1:       omp_offload.failed13:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i64 [[TMP62]], i64 [[TMP64]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
-// CHECK1:       omp_offload.cont14:
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP70]], i64* [[TMP75]], align 8
+// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP70]], i64* [[TMP77]], align 8
+// CHECK1-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP78]], align 8
+// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP72]], i64* [[TMP80]], align 8
+// CHECK1-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP72]], i64* [[TMP82]], align 8
+// CHECK1-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP83]], align 8
+// CHECK1-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS13:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP86]], align 4
+// CHECK1-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP87]], align 4
+// CHECK1-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP84]], i8*** [[TMP88]], align 8
+// CHECK1-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP85]], i8*** [[TMP89]], align 8
+// CHECK1-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP90]], align 8
+// CHECK1-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP91]], align 8
+// CHECK1-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP92]], align 8
+// CHECK1-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP93]], align 8
+// CHECK1-NEXT:    [[TMP94:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]])
+// CHECK1-NEXT:    [[TMP95:%.*]] = icmp ne i32 [[TMP94]], 0
+// CHECK1-NEXT:    br i1 [[TMP95]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
+// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i64 [[TMP70]], i64 [[TMP72]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
+// CHECK1:       omp_offload.cont15:
 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK1:       omp_if.else:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i64 [[TMP62]], i64 [[TMP64]]) #[[ATTR4]]
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i64 [[TMP70]], i64 [[TMP72]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP80:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    store i32 [[TMP80]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK1-NEXT:    [[TMP81:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    [[CONV16:%.*]] = bitcast i64* [[A_CASTED15]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP81]], i32* [[CONV16]], align 4
-// CHECK1-NEXT:    [[TMP82:%.*]] = load i64, i64* [[A_CASTED15]], align 8
-// CHECK1-NEXT:    [[TMP83:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK1-NEXT:    [[CONV17:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP83]], i32* [[CONV17]], align 4
-// CHECK1-NEXT:    [[TMP84:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK1-NEXT:    [[TMP85:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP18:%.*]] = icmp sgt i32 [[TMP85]], 20
-// CHECK1-NEXT:    br i1 [[CMP18]], label [[OMP_IF_THEN19:%.*]], label [[OMP_IF_ELSE25:%.*]]
-// CHECK1:       omp_if.then19:
-// CHECK1-NEXT:    [[TMP86:%.*]] = mul nuw i64 [[TMP2]], 4
-// CHECK1-NEXT:    [[TMP87:%.*]] = mul nuw i64 5, [[TMP5]]
-// CHECK1-NEXT:    [[TMP88:%.*]] = mul nuw i64 [[TMP87]], 8
-// CHECK1-NEXT:    [[TMP89:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP89]], i8* align 8 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i64 80, i1 false)
-// CHECK1-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP82]], i64* [[TMP91]], align 8
-// CHECK1-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP82]], i64* [[TMP93]], align 8
-// CHECK1-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP94]], align 8
-// CHECK1-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to [10 x float]**
-// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP96]], align 8
-// CHECK1-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to [10 x float]**
-// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP98]], align 8
-// CHECK1-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP99]], align 8
-// CHECK1-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP101]], align 8
-// CHECK1-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP103]], align 8
-// CHECK1-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP104]], align 8
-// CHECK1-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP106:%.*]] = bitcast i8** [[TMP105]] to float**
-// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP106]], align 8
-// CHECK1-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to float**
-// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP108]], align 8
-// CHECK1-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK1-NEXT:    store i64 [[TMP86]], i64* [[TMP109]], align 8
-// CHECK1-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 3
+// CHECK1-NEXT:    [[TMP96:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    store i32 [[TMP96]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT:    [[TMP97:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[CONV17:%.*]] = bitcast i64* [[A_CASTED16]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP97]], i32* [[CONV17]], align 4
+// CHECK1-NEXT:    [[TMP98:%.*]] = load i64, i64* [[A_CASTED16]], align 8
+// CHECK1-NEXT:    [[TMP99:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT:    [[CONV18:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP99]], i32* [[CONV18]], align 4
+// CHECK1-NEXT:    [[TMP100:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK1-NEXT:    [[TMP101:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP19:%.*]] = icmp sgt i32 [[TMP101]], 20
+// CHECK1-NEXT:    br i1 [[CMP19]], label [[OMP_IF_THEN20:%.*]], label [[OMP_IF_ELSE27:%.*]]
+// CHECK1:       omp_if.then20:
+// CHECK1-NEXT:    [[TMP102:%.*]] = mul nuw i64 [[TMP2]], 4
+// CHECK1-NEXT:    [[TMP103:%.*]] = mul nuw i64 5, [[TMP5]]
+// CHECK1-NEXT:    [[TMP104:%.*]] = mul nuw i64 [[TMP103]], 8
+// CHECK1-NEXT:    [[TMP105:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP105]], i8* align 8 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i64 80, i1 false)
+// CHECK1-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP98]], i64* [[TMP107]], align 8
+// CHECK1-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP98]], i64* [[TMP109]], align 8
+// CHECK1-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
 // CHECK1-NEXT:    store i8* null, i8** [[TMP110]], align 8
-// CHECK1-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to [5 x [10 x double]]**
-// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP112]], align 8
-// CHECK1-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to [5 x [10 x double]]**
-// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP114]], align 8
-// CHECK1-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 4
+// CHECK1-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to [10 x float]**
+// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP112]], align 8
+// CHECK1-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to [10 x float]**
+// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP114]], align 8
+// CHECK1-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
 // CHECK1-NEXT:    store i8* null, i8** [[TMP115]], align 8
-// CHECK1-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i64*
-// CHECK1-NEXT:    store i64 5, i64* [[TMP117]], align 8
-// CHECK1-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 5
+// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP117]], align 8
+// CHECK1-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP119:%.*]] = bitcast i8** [[TMP118]] to i64*
-// CHECK1-NEXT:    store i64 5, i64* [[TMP119]], align 8
-// CHECK1-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 5
+// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP119]], align 8
+// CHECK1-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
 // CHECK1-NEXT:    store i8* null, i8** [[TMP120]], align 8
-// CHECK1-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP122]], align 8
-// CHECK1-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP124]], align 8
-// CHECK1-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 6
-// CHECK1-NEXT:    store i8* null, i8** [[TMP125]], align 8
-// CHECK1-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to double**
-// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP127]], align 8
-// CHECK1-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to double**
-// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP129]], align 8
-// CHECK1-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK1-NEXT:    store i64 [[TMP88]], i64* [[TMP130]], align 8
-// CHECK1-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 7
+// CHECK1-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to float**
+// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP122]], align 8
+// CHECK1-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to float**
+// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP124]], align 8
+// CHECK1-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK1-NEXT:    store i64 [[TMP102]], i64* [[TMP125]], align 8
+// CHECK1-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK1-NEXT:    store i8* null, i8** [[TMP126]], align 8
+// CHECK1-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to [5 x [10 x double]]**
+// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP128]], align 8
+// CHECK1-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to [5 x [10 x double]]**
+// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP130]], align 8
+// CHECK1-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 4
 // CHECK1-NEXT:    store i8* null, i8** [[TMP131]], align 8
-// CHECK1-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to %struct.TT**
-// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP133]], align 8
-// CHECK1-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to %struct.TT**
-// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP135]], align 8
-// CHECK1-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 8
+// CHECK1-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i64*
+// CHECK1-NEXT:    store i64 5, i64* [[TMP133]], align 8
+// CHECK1-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i64*
+// CHECK1-NEXT:    store i64 5, i64* [[TMP135]], align 8
+// CHECK1-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 5
 // CHECK1-NEXT:    store i8* null, i8** [[TMP136]], align 8
-// CHECK1-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 9
+// CHECK1-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 6
 // CHECK1-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP84]], i64* [[TMP138]], align 8
-// CHECK1-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 9
+// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP138]], align 8
+// CHECK1-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 6
 // CHECK1-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP84]], i64* [[TMP140]], align 8
-// CHECK1-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 9
+// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP140]], align 8
+// CHECK1-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 6
 // CHECK1-NEXT:    store i8* null, i8** [[TMP141]], align 8
-// CHECK1-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP145:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170.region_id, i32 10, i8** [[TMP142]], i8** [[TMP143]], i64* [[TMP144]], i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP146:%.*]] = icmp ne i32 [[TMP145]], 0
-// CHECK1-NEXT:    br i1 [[TMP146]], label [[OMP_OFFLOAD_FAILED23:%.*]], label [[OMP_OFFLOAD_CONT24:%.*]]
-// CHECK1:       omp_offload.failed23:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i64 [[TMP82]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP84]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT24]]
-// CHECK1:       omp_offload.cont24:
-// CHECK1-NEXT:    br label [[OMP_IF_END26:%.*]]
-// CHECK1:       omp_if.else25:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i64 [[TMP82]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP84]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_IF_END26]]
-// CHECK1:       omp_if.end26:
-// CHECK1-NEXT:    [[TMP147:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    [[TMP148:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP148]])
-// CHECK1-NEXT:    ret i32 [[TMP147]]
+// CHECK1-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to double**
+// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP143]], align 8
+// CHECK1-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP145:%.*]] = bitcast i8** [[TMP144]] to double**
+// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP145]], align 8
+// CHECK1-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK1-NEXT:    store i64 [[TMP104]], i64* [[TMP146]], align 8
+// CHECK1-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 7
+// CHECK1-NEXT:    store i8* null, i8** [[TMP147]], align 8
+// CHECK1-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to %struct.TT**
+// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP149]], align 8
+// CHECK1-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to %struct.TT**
+// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP151]], align 8
+// CHECK1-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 8
+// CHECK1-NEXT:    store i8* null, i8** [[TMP152]], align 8
+// CHECK1-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 9
+// CHECK1-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP100]], i64* [[TMP154]], align 8
+// CHECK1-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 9
+// CHECK1-NEXT:    [[TMP156:%.*]] = bitcast i8** [[TMP155]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP100]], i64* [[TMP156]], align 8
+// CHECK1-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 9
+// CHECK1-NEXT:    store i8* null, i8** [[TMP157]], align 8
+// CHECK1-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS24:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP161]], align 4
+// CHECK1-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 10, i32* [[TMP162]], align 4
+// CHECK1-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP158]], i8*** [[TMP163]], align 8
+// CHECK1-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP159]], i8*** [[TMP164]], align 8
+// CHECK1-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP160]], i64** [[TMP165]], align 8
+// CHECK1-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP166]], align 8
+// CHECK1-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP167]], align 8
+// CHECK1-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP168]], align 8
+// CHECK1-NEXT:    [[TMP169:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]])
+// CHECK1-NEXT:    [[TMP170:%.*]] = icmp ne i32 [[TMP169]], 0
+// CHECK1-NEXT:    br i1 [[TMP170]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
+// CHECK1:       omp_offload.failed25:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i64 [[TMP98]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP100]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
+// CHECK1:       omp_offload.cont26:
+// CHECK1-NEXT:    br label [[OMP_IF_END28:%.*]]
+// CHECK1:       omp_if.else27:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i64 [[TMP98]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP100]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_IF_END28]]
+// CHECK1:       omp_if.end28:
+// CHECK1-NEXT:    [[TMP171:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP172:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP172]])
+// CHECK1-NEXT:    ret i32 [[TMP171]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103
@@ -1003,6 +1054,7 @@ int bar(int n){
 // CHECK1-NEXT:    [[AA_CASTED_I:%.*]] = alloca i64, align 8
 // CHECK1-NEXT:    [[LIN_CASTED_I:%.*]] = alloca i64, align 8
 // CHECK1-NEXT:    [[A_CASTED_I:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK1-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -1041,23 +1093,39 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[TMP19]], i64 0, i64 0
 // CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP12]], i32 0, i32 1
 // CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__3_EXIT:%.*]]
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4, !noalias !24
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP26]], align 4, !noalias !24
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP27]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP28]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP22]], i64** [[TMP29]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP30]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP32]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK1-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
+// CHECK1-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__3_EXIT:%.*]]
 // CHECK1:       omp_offload.failed.i:
-// CHECK1-NEXT:    [[TMP27:%.*]] = load i16, i16* [[TMP16]], align 2
+// CHECK1-NEXT:    [[TMP35:%.*]] = load i16, i16* [[TMP16]], align 2
 // CHECK1-NEXT:    [[CONV_I:%.*]] = bitcast i64* [[AA_CASTED_I]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP27]], i16* [[CONV_I]], align 2, !noalias !24
-// CHECK1-NEXT:    [[TMP28:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !24
-// CHECK1-NEXT:    [[TMP29:%.*]] = load i32, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    store i16 [[TMP35]], i16* [[CONV_I]], align 2, !noalias !24
+// CHECK1-NEXT:    [[TMP36:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP37:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK1-NEXT:    [[CONV4_I:%.*]] = bitcast i64* [[LIN_CASTED_I]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP29]], i32* [[CONV4_I]], align 4, !noalias !24
-// CHECK1-NEXT:    [[TMP30:%.*]] = load i64, i64* [[LIN_CASTED_I]], align 8, !noalias !24
-// CHECK1-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    store i32 [[TMP37]], i32* [[CONV4_I]], align 4, !noalias !24
+// CHECK1-NEXT:    [[TMP38:%.*]] = load i64, i64* [[LIN_CASTED_I]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP39:%.*]] = load i32, i32* [[TMP24]], align 4
 // CHECK1-NEXT:    [[CONV5_I:%.*]] = bitcast i64* [[A_CASTED_I]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP31]], i32* [[CONV5_I]], align 4, !noalias !24
-// CHECK1-NEXT:    [[TMP32:%.*]] = load i64, i64* [[A_CASTED_I]], align 8, !noalias !24
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138(i64 [[TMP28]], i64 [[TMP30]], i64 [[TMP32]]) #[[ATTR4]]
+// CHECK1-NEXT:    store i32 [[TMP39]], i32* [[CONV5_I]], align 4, !noalias !24
+// CHECK1-NEXT:    [[TMP40:%.*]] = load i64, i64* [[A_CASTED_I]], align 8, !noalias !24
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138(i64 [[TMP36]], i64 [[TMP38]], i64 [[TMP40]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[DOTOMP_OUTLINED__3_EXIT]]
 // CHECK1:       .omp_outlined..3.exit:
 // CHECK1-NEXT:    ret i32 0
@@ -1469,9 +1537,26 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK1-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK1-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1481,15 +1566,15 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP42:%.*]] = mul nsw i64 1, [[TMP2]]
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP42]]
+// CHECK1-NEXT:    [[TMP50:%.*]] = mul nsw i64 1, [[TMP2]]
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP50]]
 // CHECK1-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
-// CHECK1-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK1-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP44]]
-// CHECK1-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK1-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP52]]
+// CHECK1-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK1-NEXT:    ret i32 [[ADD4]]
 //
 //
@@ -1561,9 +1646,26 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
 // CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK1-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK1-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1573,8 +1675,8 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP31]]
+// CHECK1-NEXT:    [[TMP39:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP39]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -1630,9 +1732,26 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP19]], align 8
 // CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK1-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK1-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1642,8 +1761,8 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP24]]
+// CHECK1-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242
@@ -1968,11 +2087,11 @@ int bar(int n){
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS7:%.*]] = alloca [2 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS8:%.*]] = alloca [2 x i8*], align 4
 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[A_CASTED11:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[A_CASTED12:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS14:%.*]] = alloca [10 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS15:%.*]] = alloca [10 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS16:%.*]] = alloca [10 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [10 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [10 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [10 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [10 x i64], align 4
 // CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
@@ -1987,252 +2106,303 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP4:%.*]] = mul nuw i32 5, [[TMP3]]
 // CHECK3-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP4]], align 8
 // CHECK3-NEXT:    store i32 [[TMP3]], i32* [[__VLA_EXPR1]], align 4
-// CHECK3-NEXT:    [[TMP5:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP6:%.*]] = icmp ne i32 [[TMP5]], 0
-// CHECK3-NEXT:    br i1 [[TMP6]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
+// CHECK3-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103() #[[ATTR4:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[CALL:%.*]] = call noundef i64 @_Z7get_valv()
 // CHECK3-NEXT:    store i64 [[CALL]], i64* [[K]], align 8
-// CHECK3-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP7]], i32* [[A_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_CASTED]], align 4
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110(i32 [[TMP8]], i64* [[K]]) #[[ATTR4]]
+// CHECK3-NEXT:    [[TMP15:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP15]], i32* [[A_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = load i32, i32* [[A_CASTED]], align 4
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110(i32 [[TMP16]], i64* [[K]]) #[[ATTR4]]
 // CHECK3-NEXT:    store i32 12, i32* [[LIN]], align 4
-// CHECK3-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK3-NEXT:    [[TMP17:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP9]], i16* [[CONV]], align 2
-// CHECK3-NEXT:    [[TMP10:%.*]] = load i32, i32* [[AA_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = load i32, i32* [[LIN]], align 4
-// CHECK3-NEXT:    store i32 [[TMP11]], i32* [[LIN_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[LIN_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP13]], i32* [[A_CASTED2]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = load i32, i32* [[A_CASTED2]], align 4
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP10]], i32* [[TMP16]], align 4
-// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP10]], i32* [[TMP18]], align 4
-// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP21]], align 4
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP23]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP24]], align 4
-// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK3-NEXT:    store i16 [[TMP17]], i16* [[CONV]], align 2
+// CHECK3-NEXT:    [[TMP18:%.*]] = load i32, i32* [[AA_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = load i32, i32* [[LIN]], align 4
+// CHECK3-NEXT:    store i32 [[TMP19]], i32* [[LIN_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP20:%.*]] = load i32, i32* [[LIN_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP21]], i32* [[A_CASTED2]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = load i32, i32* [[A_CASTED2]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP18]], i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP14]], i32* [[TMP26]], align 4
-// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP14]], i32* [[TMP28]], align 4
-// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP29]], align 4
-// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP33:%.*]] = load i16, i16* [[AA]], align 2
-// CHECK3-NEXT:    store i16 [[TMP33]], i16* [[TMP32]], align 4
-// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP35:%.*]] = load i32, i32* [[LIN]], align 4
-// CHECK3-NEXT:    store i32 [[TMP35]], i32* [[TMP34]], align 4
-// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP37:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP37]], i32* [[TMP36]], align 4
-// CHECK3-NEXT:    [[TMP38:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i32 1, i32 72, i32 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
-// CHECK3-NEXT:    [[TMP39:%.*]] = bitcast i8* [[TMP38]] to %struct.kmp_task_t_with_privates*
-// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP39]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP40]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP42:%.*]] = load i8*, i8** [[TMP41]], align 4
-// CHECK3-NEXT:    [[TMP43:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
-// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP42]], i8* align 4 [[TMP43]], i32 12, i1 false)
-// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP39]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP45:%.*]] = bitcast i8* [[TMP42]] to %struct.anon*
-// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP47:%.*]] = bitcast [3 x i64]* [[TMP46]] to i8*
-// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP47]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes to i8*), i32 24, i1 false)
-// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP49:%.*]] = bitcast [3 x i8*]* [[TMP48]] to i8*
-// CHECK3-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP30]] to i8*
-// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP49]], i8* align 4 [[TMP50]], i32 12, i1 false)
-// CHECK3-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP52:%.*]] = bitcast [3 x i8*]* [[TMP51]] to i8*
-// CHECK3-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP31]] to i8*
-// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP52]], i8* align 4 [[TMP53]], i32 12, i1 false)
-// CHECK3-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP55:%.*]] = load i16, i16* [[AA]], align 2
-// CHECK3-NEXT:    store i16 [[TMP55]], i16* [[TMP54]], align 4
-// CHECK3-NEXT:    [[TMP56:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i8* [[TMP38]])
-// CHECK3-NEXT:    [[TMP57:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP57]], i32* [[A_CASTED3]], align 4
-// CHECK3-NEXT:    [[TMP58:%.*]] = load i32, i32* [[A_CASTED3]], align 4
-// CHECK3-NEXT:    [[TMP59:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK3-NEXT:    store i32 [[TMP18]], i32* [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP20]], i32* [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP20]], i32* [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP34:%.*]] = bitcast i8** [[TMP33]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP22]], i32* [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP22]], i32* [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8* null, i8** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP41:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK3-NEXT:    store i16 [[TMP41]], i16* [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP43:%.*]] = load i32, i32* [[LIN]], align 4
+// CHECK3-NEXT:    store i32 [[TMP43]], i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP45:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP45]], i32* [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i32 1, i32 72, i32 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
+// CHECK3-NEXT:    [[TMP47:%.*]] = bitcast i8* [[TMP46]] to %struct.kmp_task_t_with_privates*
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP47]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP48]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP50:%.*]] = load i8*, i8** [[TMP49]], align 4
+// CHECK3-NEXT:    [[TMP51:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
+// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP50]], i8* align 4 [[TMP51]], i32 12, i1 false)
+// CHECK3-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP47]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP53:%.*]] = bitcast i8* [[TMP50]] to %struct.anon*
+// CHECK3-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP52]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP55:%.*]] = bitcast [3 x i64]* [[TMP54]] to i8*
+// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP55]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes to i8*), i32 24, i1 false)
+// CHECK3-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP52]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP57:%.*]] = bitcast [3 x i8*]* [[TMP56]] to i8*
+// CHECK3-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP38]] to i8*
+// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP57]], i8* align 4 [[TMP58]], i32 12, i1 false)
+// CHECK3-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP52]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP60:%.*]] = bitcast [3 x i8*]* [[TMP59]] to i8*
+// CHECK3-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP39]] to i8*
+// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP60]], i8* align 4 [[TMP61]], i32 12, i1 false)
+// CHECK3-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP52]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP63:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK3-NEXT:    store i16 [[TMP63]], i16* [[TMP62]], align 4
+// CHECK3-NEXT:    [[TMP64:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i8* [[TMP46]])
+// CHECK3-NEXT:    [[TMP65:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP65]], i32* [[A_CASTED3]], align 4
+// CHECK3-NEXT:    [[TMP66:%.*]] = load i32, i32* [[A_CASTED3]], align 4
+// CHECK3-NEXT:    [[TMP67:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK3-NEXT:    [[CONV5:%.*]] = bitcast i32* [[AA_CASTED4]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP59]], i16* [[CONV5]], align 2
-// CHECK3-NEXT:    [[TMP60:%.*]] = load i32, i32* [[AA_CASTED4]], align 4
-// CHECK3-NEXT:    [[TMP61:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP61]], 10
+// CHECK3-NEXT:    store i16 [[TMP67]], i16* [[CONV5]], align 2
+// CHECK3-NEXT:    [[TMP68:%.*]] = load i32, i32* [[AA_CASTED4]], align 4
+// CHECK3-NEXT:    [[TMP69:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP69]], 10
 // CHECK3-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK3:       omp_if.then:
-// CHECK3-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP63:%.*]] = bitcast i8** [[TMP62]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP58]], i32* [[TMP63]], align 4
-// CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP65:%.*]] = bitcast i8** [[TMP64]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP58]], i32* [[TMP65]], align 4
-// CHECK3-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP66]], align 4
-// CHECK3-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP60]], i32* [[TMP68]], align 4
-// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP60]], i32* [[TMP70]], align 4
-// CHECK3-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP71]], align 4
-// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146.region_id, i32 2, i8** [[TMP72]], i8** [[TMP73]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
-// CHECK3-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED9:%.*]], label [[OMP_OFFLOAD_CONT10:%.*]]
-// CHECK3:       omp_offload.failed9:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i32 [[TMP58]], i32 [[TMP60]]) #[[ATTR4]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT10]]
-// CHECK3:       omp_offload.cont10:
+// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP66]], i32* [[TMP71]], align 4
+// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP66]], i32* [[TMP73]], align 4
+// CHECK3-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP74]], align 4
+// CHECK3-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP68]], i32* [[TMP76]], align 4
+// CHECK3-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP68]], i32* [[TMP78]], align 4
+// CHECK3-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP79]], align 4
+// CHECK3-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS9:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP82]], align 4
+// CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP83]], align 4
+// CHECK3-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP80]], i8*** [[TMP84]], align 4
+// CHECK3-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP81]], i8*** [[TMP85]], align 4
+// CHECK3-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP86]], align 4
+// CHECK3-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP87]], align 4
+// CHECK3-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP88]], align 4
+// CHECK3-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP89]], align 4
+// CHECK3-NEXT:    [[TMP90:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]])
+// CHECK3-NEXT:    [[TMP91:%.*]] = icmp ne i32 [[TMP90]], 0
+// CHECK3-NEXT:    br i1 [[TMP91]], label [[OMP_OFFLOAD_FAILED10:%.*]], label [[OMP_OFFLOAD_CONT11:%.*]]
+// CHECK3:       omp_offload.failed10:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i32 [[TMP66]], i32 [[TMP68]]) #[[ATTR4]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT11]]
+// CHECK3:       omp_offload.cont11:
 // CHECK3-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK3:       omp_if.else:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i32 [[TMP58]], i32 [[TMP60]]) #[[ATTR4]]
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i32 [[TMP66]], i32 [[TMP68]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP76:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP76]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK3-NEXT:    [[TMP77:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP77]], i32* [[A_CASTED11]], align 4
-// CHECK3-NEXT:    [[TMP78:%.*]] = load i32, i32* [[A_CASTED11]], align 4
-// CHECK3-NEXT:    [[TMP79:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK3-NEXT:    store i32 [[TMP79]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK3-NEXT:    [[TMP80:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK3-NEXT:    [[TMP81:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP12:%.*]] = icmp sgt i32 [[TMP81]], 20
-// CHECK3-NEXT:    br i1 [[CMP12]], label [[OMP_IF_THEN13:%.*]], label [[OMP_IF_ELSE19:%.*]]
-// CHECK3:       omp_if.then13:
-// CHECK3-NEXT:    [[TMP82:%.*]] = mul nuw i32 [[TMP1]], 4
-// CHECK3-NEXT:    [[TMP83:%.*]] = sext i32 [[TMP82]] to i64
-// CHECK3-NEXT:    [[TMP84:%.*]] = mul nuw i32 5, [[TMP3]]
-// CHECK3-NEXT:    [[TMP85:%.*]] = mul nuw i32 [[TMP84]], 8
-// CHECK3-NEXT:    [[TMP86:%.*]] = sext i32 [[TMP85]] to i64
-// CHECK3-NEXT:    [[TMP87:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP87]], i8* align 4 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i32 80, i1 false)
-// CHECK3-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP78]], i32* [[TMP89]], align 4
-// CHECK3-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP78]], i32* [[TMP91]], align 4
-// CHECK3-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP92]], align 4
-// CHECK3-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to [10 x float]**
-// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP94]], align 4
-// CHECK3-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to [10 x float]**
-// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP96]], align 4
-// CHECK3-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP97]], align 4
-// CHECK3-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP99]], align 4
-// CHECK3-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP101]], align 4
-// CHECK3-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP102]], align 4
-// CHECK3-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to float**
-// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP104]], align 4
-// CHECK3-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP106:%.*]] = bitcast i8** [[TMP105]] to float**
-// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP106]], align 4
-// CHECK3-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK3-NEXT:    store i64 [[TMP83]], i64* [[TMP107]], align 4
-// CHECK3-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP92:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP92]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK3-NEXT:    [[TMP93:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP93]], i32* [[A_CASTED12]], align 4
+// CHECK3-NEXT:    [[TMP94:%.*]] = load i32, i32* [[A_CASTED12]], align 4
+// CHECK3-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK3-NEXT:    store i32 [[TMP95]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK3-NEXT:    [[TMP96:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK3-NEXT:    [[TMP97:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP13:%.*]] = icmp sgt i32 [[TMP97]], 20
+// CHECK3-NEXT:    br i1 [[CMP13]], label [[OMP_IF_THEN14:%.*]], label [[OMP_IF_ELSE21:%.*]]
+// CHECK3:       omp_if.then14:
+// CHECK3-NEXT:    [[TMP98:%.*]] = mul nuw i32 [[TMP1]], 4
+// CHECK3-NEXT:    [[TMP99:%.*]] = sext i32 [[TMP98]] to i64
+// CHECK3-NEXT:    [[TMP100:%.*]] = mul nuw i32 5, [[TMP3]]
+// CHECK3-NEXT:    [[TMP101:%.*]] = mul nuw i32 [[TMP100]], 8
+// CHECK3-NEXT:    [[TMP102:%.*]] = sext i32 [[TMP101]] to i64
+// CHECK3-NEXT:    [[TMP103:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP103]], i8* align 4 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i32 80, i1 false)
+// CHECK3-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP94]], i32* [[TMP105]], align 4
+// CHECK3-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP94]], i32* [[TMP107]], align 4
+// CHECK3-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 0
 // CHECK3-NEXT:    store i8* null, i8** [[TMP108]], align 4
-// CHECK3-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to [5 x [10 x double]]**
-// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP110]], align 4
-// CHECK3-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to [5 x [10 x double]]**
-// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP112]], align 4
-// CHECK3-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to [10 x float]**
+// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP110]], align 4
+// CHECK3-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to [10 x float]**
+// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP112]], align 4
+// CHECK3-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 1
 // CHECK3-NEXT:    store i8* null, i8** [[TMP113]], align 4
-// CHECK3-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32*
-// CHECK3-NEXT:    store i32 5, i32* [[TMP115]], align 4
-// CHECK3-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP115]], align 4
+// CHECK3-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i32*
-// CHECK3-NEXT:    store i32 5, i32* [[TMP117]], align 4
-// CHECK3-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 5
+// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP117]], align 4
+// CHECK3-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 2
 // CHECK3-NEXT:    store i8* null, i8** [[TMP118]], align 4
-// CHECK3-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 6
-// CHECK3-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP120]], align 4
-// CHECK3-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 6
-// CHECK3-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP122]], align 4
-// CHECK3-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 6
-// CHECK3-NEXT:    store i8* null, i8** [[TMP123]], align 4
-// CHECK3-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 7
-// CHECK3-NEXT:    [[TMP125:%.*]] = bitcast i8** [[TMP124]] to double**
-// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP125]], align 4
-// CHECK3-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 7
-// CHECK3-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to double**
-// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP127]], align 4
-// CHECK3-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK3-NEXT:    store i64 [[TMP86]], i64* [[TMP128]], align 4
-// CHECK3-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to float**
+// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP120]], align 4
+// CHECK3-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to float**
+// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP122]], align 4
+// CHECK3-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK3-NEXT:    store i64 [[TMP99]], i64* [[TMP123]], align 4
+// CHECK3-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 3
+// CHECK3-NEXT:    store i8* null, i8** [[TMP124]], align 4
+// CHECK3-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to [5 x [10 x double]]**
+// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP126]], align 4
+// CHECK3-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to [5 x [10 x double]]**
+// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP128]], align 4
+// CHECK3-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 4
 // CHECK3-NEXT:    store i8* null, i8** [[TMP129]], align 4
-// CHECK3-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 8
-// CHECK3-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to %struct.TT**
-// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP131]], align 4
-// CHECK3-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 8
-// CHECK3-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to %struct.TT**
-// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP133]], align 4
-// CHECK3-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i32*
+// CHECK3-NEXT:    store i32 5, i32* [[TMP131]], align 4
+// CHECK3-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i32*
+// CHECK3-NEXT:    store i32 5, i32* [[TMP133]], align 4
+// CHECK3-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 5
 // CHECK3-NEXT:    store i8* null, i8** [[TMP134]], align 4
-// CHECK3-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 9
+// CHECK3-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 6
 // CHECK3-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP80]], i32* [[TMP136]], align 4
-// CHECK3-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 9
+// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP136]], align 4
+// CHECK3-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 6
 // CHECK3-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP80]], i32* [[TMP138]], align 4
-// CHECK3-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 9
+// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP138]], align 4
+// CHECK3-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 6
 // CHECK3-NEXT:    store i8* null, i8** [[TMP139]], align 4
-// CHECK3-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP143:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170.region_id, i32 10, i8** [[TMP140]], i8** [[TMP141]], i64* [[TMP142]], i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP144:%.*]] = icmp ne i32 [[TMP143]], 0
-// CHECK3-NEXT:    br i1 [[TMP144]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK3:       omp_offload.failed17:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i32 [[TMP78]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP80]]) #[[ATTR4]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK3:       omp_offload.cont18:
-// CHECK3-NEXT:    br label [[OMP_IF_END20:%.*]]
-// CHECK3:       omp_if.else19:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i32 [[TMP78]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP80]]) #[[ATTR4]]
-// CHECK3-NEXT:    br label [[OMP_IF_END20]]
-// CHECK3:       omp_if.end20:
-// CHECK3-NEXT:    [[TMP145:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    [[TMP146:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP146]])
-// CHECK3-NEXT:    ret i32 [[TMP145]]
+// CHECK3-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to double**
+// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP141]], align 4
+// CHECK3-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to double**
+// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP143]], align 4
+// CHECK3-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK3-NEXT:    store i64 [[TMP102]], i64* [[TMP144]], align 4
+// CHECK3-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 7
+// CHECK3-NEXT:    store i8* null, i8** [[TMP145]], align 4
+// CHECK3-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to %struct.TT**
+// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP147]], align 4
+// CHECK3-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to %struct.TT**
+// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP149]], align 4
+// CHECK3-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 8
+// CHECK3-NEXT:    store i8* null, i8** [[TMP150]], align 4
+// CHECK3-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 9
+// CHECK3-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP96]], i32* [[TMP152]], align 4
+// CHECK3-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 9
+// CHECK3-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP96]], i32* [[TMP154]], align 4
+// CHECK3-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 9
+// CHECK3-NEXT:    store i8* null, i8** [[TMP155]], align 4
+// CHECK3-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS18:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP159]], align 4
+// CHECK3-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 10, i32* [[TMP160]], align 4
+// CHECK3-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP156]], i8*** [[TMP161]], align 4
+// CHECK3-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP157]], i8*** [[TMP162]], align 4
+// CHECK3-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP158]], i64** [[TMP163]], align 4
+// CHECK3-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP164]], align 4
+// CHECK3-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP165]], align 4
+// CHECK3-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP166]], align 4
+// CHECK3-NEXT:    [[TMP167:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]])
+// CHECK3-NEXT:    [[TMP168:%.*]] = icmp ne i32 [[TMP167]], 0
+// CHECK3-NEXT:    br i1 [[TMP168]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
+// CHECK3:       omp_offload.failed19:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i32 [[TMP94]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP96]]) #[[ATTR4]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
+// CHECK3:       omp_offload.cont20:
+// CHECK3-NEXT:    br label [[OMP_IF_END22:%.*]]
+// CHECK3:       omp_if.else21:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i32 [[TMP94]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP96]]) #[[ATTR4]]
+// CHECK3-NEXT:    br label [[OMP_IF_END22]]
+// CHECK3:       omp_if.end22:
+// CHECK3-NEXT:    [[TMP169:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    [[TMP170:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP170]])
+// CHECK3-NEXT:    ret i32 [[TMP169]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103
@@ -2600,6 +2770,7 @@ int bar(int n){
 // CHECK3-NEXT:    [[AA_CASTED_I:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[LIN_CASTED_I:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[A_CASTED_I:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK3-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 4
 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -2638,21 +2809,37 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[TMP19]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP12]], i32 0, i32 1
 // CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK3-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK3-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__3_EXIT:%.*]]
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP25]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP26]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP20]], i8*** [[TMP27]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP28]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP22]], i64** [[TMP29]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP30]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP32]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK3-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
+// CHECK3-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__3_EXIT:%.*]]
 // CHECK3:       omp_offload.failed.i:
-// CHECK3-NEXT:    [[TMP27:%.*]] = load i16, i16* [[TMP16]], align 2
+// CHECK3-NEXT:    [[TMP35:%.*]] = load i16, i16* [[TMP16]], align 2
 // CHECK3-NEXT:    [[CONV_I:%.*]] = bitcast i32* [[AA_CASTED_I]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP27]], i16* [[CONV_I]], align 2, !noalias !25
-// CHECK3-NEXT:    [[TMP28:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !25
-// CHECK3-NEXT:    [[TMP29:%.*]] = load i32, i32* [[TMP23]], align 4
-// CHECK3-NEXT:    store i32 [[TMP29]], i32* [[LIN_CASTED_I]], align 4, !noalias !25
-// CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[LIN_CASTED_I]], align 4, !noalias !25
-// CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP24]], align 4
-// CHECK3-NEXT:    store i32 [[TMP31]], i32* [[A_CASTED_I]], align 4, !noalias !25
-// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A_CASTED_I]], align 4, !noalias !25
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138(i32 [[TMP28]], i32 [[TMP30]], i32 [[TMP32]]) #[[ATTR4]]
+// CHECK3-NEXT:    store i16 [[TMP35]], i16* [[CONV_I]], align 2, !noalias !25
+// CHECK3-NEXT:    [[TMP36:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP37:%.*]] = load i32, i32* [[TMP23]], align 4
+// CHECK3-NEXT:    store i32 [[TMP37]], i32* [[LIN_CASTED_I]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP38:%.*]] = load i32, i32* [[LIN_CASTED_I]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP39:%.*]] = load i32, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    store i32 [[TMP39]], i32* [[A_CASTED_I]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP40:%.*]] = load i32, i32* [[A_CASTED_I]], align 4, !noalias !25
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138(i32 [[TMP36]], i32 [[TMP38]], i32 [[TMP40]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[DOTOMP_OUTLINED__3_EXIT]]
 // CHECK3:       .omp_outlined..3.exit:
 // CHECK3-NEXT:    ret i32 0
@@ -3054,9 +3241,26 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK3-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK3-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3066,15 +3270,15 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP42:%.*]] = mul nsw i32 1, [[TMP1]]
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP42]]
+// CHECK3-NEXT:    [[TMP50:%.*]] = mul nsw i32 1, [[TMP1]]
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP50]]
 // CHECK3-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
-// CHECK3-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK3-NEXT:    [[CONV:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK3-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP44]]
-// CHECK3-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK3-NEXT:    [[CONV:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK3-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP52]]
+// CHECK3-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK3-NEXT:    ret i32 [[ADD3]]
 //
 //
@@ -3145,9 +3349,26 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP26]], align 4
 // CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK3-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK3-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3157,8 +3378,8 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP31]]
+// CHECK3-NEXT:    [[TMP39:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP39]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -3213,9 +3434,26 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP19]], align 4
 // CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK3-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK3-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3225,8 +3463,8 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP24]]
+// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242
@@ -5127,11 +5365,11 @@ int bar(int n){
 // CHECK17-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [2 x i8*], align 8
 // CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [2 x i8*], align 8
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[A_CASTED15:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[A_CASTED16:%.*]] = alloca i64, align 8
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [10 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [10 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [10 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [10 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [10 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [10 x i8*], align 8
 // CHECK17-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [10 x i64], align 8
 // CHECK17-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK17-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
@@ -5148,259 +5386,310 @@ int bar(int n){
 // CHECK17-NEXT:    [[TMP6:%.*]] = mul nuw i64 5, [[TMP5]]
 // CHECK17-NEXT:    [[VLA1:%.*]] = alloca double, i64 [[TMP6]], align 8
 // CHECK17-NEXT:    store i64 [[TMP5]], i64* [[__VLA_EXPR1]], align 8
-// CHECK17-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0)
-// CHECK17-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK17-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK17-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 0, i32* [[TMP8]], align 4
+// CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP9]], align 8
+// CHECK17-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP10]], align 8
+// CHECK17-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* null, i64** [[TMP11]], align 8
+// CHECK17-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* null, i64** [[TMP12]], align 8
+// CHECK17-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK17-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK17-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103() #[[ATTR4:[0-9]+]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
 // CHECK17-NEXT:    [[CALL:%.*]] = call noundef i64 @_Z7get_valv()
 // CHECK17-NEXT:    store i64 [[CALL]], i64* [[K]], align 8
-// CHECK17-NEXT:    [[TMP9:%.*]] = load i32, i32* [[A]], align 4
+// CHECK17-NEXT:    [[TMP17:%.*]] = load i32, i32* [[A]], align 4
 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP9]], i32* [[CONV]], align 4
-// CHECK17-NEXT:    [[TMP10:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK17-NEXT:    [[TMP11:%.*]] = load i64, i64* [[K]], align 8
-// CHECK17-NEXT:    store i64 [[TMP11]], i64* [[K_CASTED]], align 8
-// CHECK17-NEXT:    [[TMP12:%.*]] = load i64, i64* [[K_CASTED]], align 8
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110(i64 [[TMP10]], i64 [[TMP12]]) #[[ATTR4]]
+// CHECK17-NEXT:    store i32 [[TMP17]], i32* [[CONV]], align 4
+// CHECK17-NEXT:    [[TMP18:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK17-NEXT:    [[TMP19:%.*]] = load i64, i64* [[K]], align 8
+// CHECK17-NEXT:    store i64 [[TMP19]], i64* [[K_CASTED]], align 8
+// CHECK17-NEXT:    [[TMP20:%.*]] = load i64, i64* [[K_CASTED]], align 8
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110(i64 [[TMP18]], i64 [[TMP20]]) #[[ATTR4]]
 // CHECK17-NEXT:    store i32 12, i32* [[LIN]], align 4
-// CHECK17-NEXT:    [[TMP13:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK17-NEXT:    [[TMP21:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK17-NEXT:    [[CONV2:%.*]] = bitcast i64* [[AA_CASTED]] to i16*
-// CHECK17-NEXT:    store i16 [[TMP13]], i16* [[CONV2]], align 2
-// CHECK17-NEXT:    [[TMP14:%.*]] = load i64, i64* [[AA_CASTED]], align 8
-// CHECK17-NEXT:    [[TMP15:%.*]] = load i32, i32* [[LIN]], align 4
+// CHECK17-NEXT:    store i16 [[TMP21]], i16* [[CONV2]], align 2
+// CHECK17-NEXT:    [[TMP22:%.*]] = load i64, i64* [[AA_CASTED]], align 8
+// CHECK17-NEXT:    [[TMP23:%.*]] = load i32, i32* [[LIN]], align 4
 // CHECK17-NEXT:    [[CONV3:%.*]] = bitcast i64* [[LIN_CASTED]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP15]], i32* [[CONV3]], align 4
-// CHECK17-NEXT:    [[TMP16:%.*]] = load i64, i64* [[LIN_CASTED]], align 8
-// CHECK17-NEXT:    [[TMP17:%.*]] = load i32, i32* [[A]], align 4
+// CHECK17-NEXT:    store i32 [[TMP23]], i32* [[CONV3]], align 4
+// CHECK17-NEXT:    [[TMP24:%.*]] = load i64, i64* [[LIN_CASTED]], align 8
+// CHECK17-NEXT:    [[TMP25:%.*]] = load i32, i32* [[A]], align 4
 // CHECK17-NEXT:    [[CONV5:%.*]] = bitcast i64* [[A_CASTED4]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP17]], i32* [[CONV5]], align 4
-// CHECK17-NEXT:    [[TMP18:%.*]] = load i64, i64* [[A_CASTED4]], align 8
-// CHECK17-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP14]], i64* [[TMP20]], align 8
-// CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP14]], i64* [[TMP22]], align 8
-// CHECK17-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP23]], align 8
-// CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP16]], i64* [[TMP25]], align 8
-// CHECK17-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP16]], i64* [[TMP27]], align 8
-// CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP28]], align 8
-// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK17-NEXT:    store i32 [[TMP25]], i32* [[CONV5]], align 4
+// CHECK17-NEXT:    [[TMP26:%.*]] = load i64, i64* [[A_CASTED4]], align 8
+// CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP22]], i64* [[TMP28]], align 8
+// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP18]], i64* [[TMP30]], align 8
-// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP32:%.*]] = bitcast i8** [[TMP31]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP18]], i64* [[TMP32]], align 8
-// CHECK17-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP33]], align 8
-// CHECK17-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP37:%.*]] = load i16, i16* [[AA]], align 2
-// CHECK17-NEXT:    store i16 [[TMP37]], i16* [[TMP36]], align 4
-// CHECK17-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP39:%.*]] = load i32, i32* [[LIN]], align 4
-// CHECK17-NEXT:    store i32 [[TMP39]], i32* [[TMP38]], align 4
-// CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP41:%.*]] = load i32, i32* [[A]], align 4
-// CHECK17-NEXT:    store i32 [[TMP41]], i32* [[TMP40]], align 4
-// CHECK17-NEXT:    [[TMP42:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i32 1, i64 120, i64 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
-// CHECK17-NEXT:    [[TMP43:%.*]] = bitcast i8* [[TMP42]] to %struct.kmp_task_t_with_privates*
-// CHECK17-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP43]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP44]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP46:%.*]] = load i8*, i8** [[TMP45]], align 8
-// CHECK17-NEXT:    [[TMP47:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP46]], i8* align 4 [[TMP47]], i64 12, i1 false)
-// CHECK17-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP43]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP49:%.*]] = bitcast i8* [[TMP46]] to %struct.anon*
-// CHECK17-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP48]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP51:%.*]] = bitcast [3 x i8*]* [[TMP50]] to i8*
-// CHECK17-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP34]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP51]], i8* align 8 [[TMP52]], i64 24, i1 false)
-// CHECK17-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP48]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP54:%.*]] = bitcast [3 x i8*]* [[TMP53]] to i8*
-// CHECK17-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP35]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP54]], i8* align 8 [[TMP55]], i64 24, i1 false)
-// CHECK17-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP48]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP57:%.*]] = bitcast [3 x i64]* [[TMP56]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP57]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes to i8*), i64 24, i1 false)
-// CHECK17-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP48]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP59:%.*]] = load i16, i16* [[AA]], align 2
-// CHECK17-NEXT:    store i16 [[TMP59]], i16* [[TMP58]], align 8
-// CHECK17-NEXT:    [[TMP60:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i8* [[TMP42]])
-// CHECK17-NEXT:    [[TMP61:%.*]] = load i32, i32* [[A]], align 4
+// CHECK17-NEXT:    store i64 [[TMP22]], i64* [[TMP30]], align 8
+// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP31]], align 8
+// CHECK17-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP33:%.*]] = bitcast i8** [[TMP32]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP24]], i64* [[TMP33]], align 8
+// CHECK17-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP24]], i64* [[TMP35]], align 8
+// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP36]], align 8
+// CHECK17-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP26]], i64* [[TMP38]], align 8
+// CHECK17-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP26]], i64* [[TMP40]], align 8
+// CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP41]], align 8
+// CHECK17-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP45:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK17-NEXT:    store i16 [[TMP45]], i16* [[TMP44]], align 4
+// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP47:%.*]] = load i32, i32* [[LIN]], align 4
+// CHECK17-NEXT:    store i32 [[TMP47]], i32* [[TMP46]], align 4
+// CHECK17-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP49:%.*]] = load i32, i32* [[A]], align 4
+// CHECK17-NEXT:    store i32 [[TMP49]], i32* [[TMP48]], align 4
+// CHECK17-NEXT:    [[TMP50:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i32 1, i64 120, i64 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
+// CHECK17-NEXT:    [[TMP51:%.*]] = bitcast i8* [[TMP50]] to %struct.kmp_task_t_with_privates*
+// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP51]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP52]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP54:%.*]] = load i8*, i8** [[TMP53]], align 8
+// CHECK17-NEXT:    [[TMP55:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP54]], i8* align 4 [[TMP55]], i64 12, i1 false)
+// CHECK17-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP51]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP57:%.*]] = bitcast i8* [[TMP54]] to %struct.anon*
+// CHECK17-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP56]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP59:%.*]] = bitcast [3 x i8*]* [[TMP58]] to i8*
+// CHECK17-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP42]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP59]], i8* align 8 [[TMP60]], i64 24, i1 false)
+// CHECK17-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP56]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP62:%.*]] = bitcast [3 x i8*]* [[TMP61]] to i8*
+// CHECK17-NEXT:    [[TMP63:%.*]] = bitcast i8** [[TMP43]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP62]], i8* align 8 [[TMP63]], i64 24, i1 false)
+// CHECK17-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP56]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP65:%.*]] = bitcast [3 x i64]* [[TMP64]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP65]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes to i8*), i64 24, i1 false)
+// CHECK17-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP56]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP67:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK17-NEXT:    store i16 [[TMP67]], i16* [[TMP66]], align 8
+// CHECK17-NEXT:    [[TMP68:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i8* [[TMP50]])
+// CHECK17-NEXT:    [[TMP69:%.*]] = load i32, i32* [[A]], align 4
 // CHECK17-NEXT:    [[CONV7:%.*]] = bitcast i64* [[A_CASTED6]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP61]], i32* [[CONV7]], align 4
-// CHECK17-NEXT:    [[TMP62:%.*]] = load i64, i64* [[A_CASTED6]], align 8
-// CHECK17-NEXT:    [[TMP63:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK17-NEXT:    store i32 [[TMP69]], i32* [[CONV7]], align 4
+// CHECK17-NEXT:    [[TMP70:%.*]] = load i64, i64* [[A_CASTED6]], align 8
+// CHECK17-NEXT:    [[TMP71:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK17-NEXT:    [[CONV9:%.*]] = bitcast i64* [[AA_CASTED8]] to i16*
-// CHECK17-NEXT:    store i16 [[TMP63]], i16* [[CONV9]], align 2
-// CHECK17-NEXT:    [[TMP64:%.*]] = load i64, i64* [[AA_CASTED8]], align 8
-// CHECK17-NEXT:    [[TMP65:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK17-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP65]], 10
+// CHECK17-NEXT:    store i16 [[TMP71]], i16* [[CONV9]], align 2
+// CHECK17-NEXT:    [[TMP72:%.*]] = load i64, i64* [[AA_CASTED8]], align 8
+// CHECK17-NEXT:    [[TMP73:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK17-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP73]], 10
 // CHECK17-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK17:       omp_if.then:
-// CHECK17-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP67:%.*]] = bitcast i8** [[TMP66]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP62]], i64* [[TMP67]], align 8
-// CHECK17-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP62]], i64* [[TMP69]], align 8
-// CHECK17-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP70]], align 8
-// CHECK17-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP64]], i64* [[TMP72]], align 8
-// CHECK17-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP64]], i64* [[TMP74]], align 8
-// CHECK17-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP75]], align 8
-// CHECK17-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP78:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146.region_id, i32 2, i8** [[TMP76]], i8** [[TMP77]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK17-NEXT:    [[TMP79:%.*]] = icmp ne i32 [[TMP78]], 0
-// CHECK17-NEXT:    br i1 [[TMP79]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
-// CHECK17:       omp_offload.failed13:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i64 [[TMP62]], i64 [[TMP64]]) #[[ATTR4]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
-// CHECK17:       omp_offload.cont14:
+// CHECK17-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP70]], i64* [[TMP75]], align 8
+// CHECK17-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP70]], i64* [[TMP77]], align 8
+// CHECK17-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP78]], align 8
+// CHECK17-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP72]], i64* [[TMP80]], align 8
+// CHECK17-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP72]], i64* [[TMP82]], align 8
+// CHECK17-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP83]], align 8
+// CHECK17-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK17-NEXT:    [[KERNEL_ARGS13:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP86]], align 4
+// CHECK17-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 2, i32* [[TMP87]], align 4
+// CHECK17-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP84]], i8*** [[TMP88]], align 8
+// CHECK17-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP85]], i8*** [[TMP89]], align 8
+// CHECK17-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP90]], align 8
+// CHECK17-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP91]], align 8
+// CHECK17-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP92]], align 8
+// CHECK17-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP93]], align 8
+// CHECK17-NEXT:    [[TMP94:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]])
+// CHECK17-NEXT:    [[TMP95:%.*]] = icmp ne i32 [[TMP94]], 0
+// CHECK17-NEXT:    br i1 [[TMP95]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
+// CHECK17:       omp_offload.failed14:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i64 [[TMP70]], i64 [[TMP72]]) #[[ATTR4]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
+// CHECK17:       omp_offload.cont15:
 // CHECK17-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK17:       omp_if.else:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i64 [[TMP62]], i64 [[TMP64]]) #[[ATTR4]]
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i64 [[TMP70]], i64 [[TMP72]]) #[[ATTR4]]
 // CHECK17-NEXT:    br label [[OMP_IF_END]]
 // CHECK17:       omp_if.end:
-// CHECK17-NEXT:    [[TMP80:%.*]] = load i32, i32* [[A]], align 4
-// CHECK17-NEXT:    store i32 [[TMP80]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK17-NEXT:    [[TMP81:%.*]] = load i32, i32* [[A]], align 4
-// CHECK17-NEXT:    [[CONV16:%.*]] = bitcast i64* [[A_CASTED15]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP81]], i32* [[CONV16]], align 4
-// CHECK17-NEXT:    [[TMP82:%.*]] = load i64, i64* [[A_CASTED15]], align 8
-// CHECK17-NEXT:    [[TMP83:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK17-NEXT:    [[CONV17:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP83]], i32* [[CONV17]], align 4
-// CHECK17-NEXT:    [[TMP84:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK17-NEXT:    [[TMP85:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK17-NEXT:    [[CMP18:%.*]] = icmp sgt i32 [[TMP85]], 20
-// CHECK17-NEXT:    br i1 [[CMP18]], label [[OMP_IF_THEN19:%.*]], label [[OMP_IF_ELSE25:%.*]]
-// CHECK17:       omp_if.then19:
-// CHECK17-NEXT:    [[TMP86:%.*]] = mul nuw i64 [[TMP2]], 4
-// CHECK17-NEXT:    [[TMP87:%.*]] = mul nuw i64 5, [[TMP5]]
-// CHECK17-NEXT:    [[TMP88:%.*]] = mul nuw i64 [[TMP87]], 8
-// CHECK17-NEXT:    [[TMP89:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP89]], i8* align 8 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i64 80, i1 false)
-// CHECK17-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP82]], i64* [[TMP91]], align 8
-// CHECK17-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP82]], i64* [[TMP93]], align 8
-// CHECK17-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP94]], align 8
-// CHECK17-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to [10 x float]**
-// CHECK17-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP96]], align 8
-// CHECK17-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to [10 x float]**
-// CHECK17-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP98]], align 8
-// CHECK17-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP99]], align 8
-// CHECK17-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP2]], i64* [[TMP101]], align 8
-// CHECK17-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP2]], i64* [[TMP103]], align 8
-// CHECK17-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP104]], align 8
-// CHECK17-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP106:%.*]] = bitcast i8** [[TMP105]] to float**
-// CHECK17-NEXT:    store float* [[VLA]], float** [[TMP106]], align 8
-// CHECK17-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to float**
-// CHECK17-NEXT:    store float* [[VLA]], float** [[TMP108]], align 8
-// CHECK17-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK17-NEXT:    store i64 [[TMP86]], i64* [[TMP109]], align 8
-// CHECK17-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 3
+// CHECK17-NEXT:    [[TMP96:%.*]] = load i32, i32* [[A]], align 4
+// CHECK17-NEXT:    store i32 [[TMP96]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK17-NEXT:    [[TMP97:%.*]] = load i32, i32* [[A]], align 4
+// CHECK17-NEXT:    [[CONV17:%.*]] = bitcast i64* [[A_CASTED16]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP97]], i32* [[CONV17]], align 4
+// CHECK17-NEXT:    [[TMP98:%.*]] = load i64, i64* [[A_CASTED16]], align 8
+// CHECK17-NEXT:    [[TMP99:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK17-NEXT:    [[CONV18:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP99]], i32* [[CONV18]], align 4
+// CHECK17-NEXT:    [[TMP100:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK17-NEXT:    [[TMP101:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK17-NEXT:    [[CMP19:%.*]] = icmp sgt i32 [[TMP101]], 20
+// CHECK17-NEXT:    br i1 [[CMP19]], label [[OMP_IF_THEN20:%.*]], label [[OMP_IF_ELSE27:%.*]]
+// CHECK17:       omp_if.then20:
+// CHECK17-NEXT:    [[TMP102:%.*]] = mul nuw i64 [[TMP2]], 4
+// CHECK17-NEXT:    [[TMP103:%.*]] = mul nuw i64 5, [[TMP5]]
+// CHECK17-NEXT:    [[TMP104:%.*]] = mul nuw i64 [[TMP103]], 8
+// CHECK17-NEXT:    [[TMP105:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP105]], i8* align 8 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i64 80, i1 false)
+// CHECK17-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP98]], i64* [[TMP107]], align 8
+// CHECK17-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP98]], i64* [[TMP109]], align 8
+// CHECK17-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
 // CHECK17-NEXT:    store i8* null, i8** [[TMP110]], align 8
-// CHECK17-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 4
-// CHECK17-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to [5 x [10 x double]]**
-// CHECK17-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP112]], align 8
-// CHECK17-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 4
-// CHECK17-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to [5 x [10 x double]]**
-// CHECK17-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP114]], align 8
-// CHECK17-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 4
+// CHECK17-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to [10 x float]**
+// CHECK17-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP112]], align 8
+// CHECK17-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to [10 x float]**
+// CHECK17-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP114]], align 8
+// CHECK17-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
 // CHECK17-NEXT:    store i8* null, i8** [[TMP115]], align 8
-// CHECK17-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 5
+// CHECK17-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
 // CHECK17-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i64*
-// CHECK17-NEXT:    store i64 5, i64* [[TMP117]], align 8
-// CHECK17-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 5
+// CHECK17-NEXT:    store i64 [[TMP2]], i64* [[TMP117]], align 8
+// CHECK17-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
 // CHECK17-NEXT:    [[TMP119:%.*]] = bitcast i8** [[TMP118]] to i64*
-// CHECK17-NEXT:    store i64 5, i64* [[TMP119]], align 8
-// CHECK17-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 5
+// CHECK17-NEXT:    store i64 [[TMP2]], i64* [[TMP119]], align 8
+// CHECK17-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
 // CHECK17-NEXT:    store i8* null, i8** [[TMP120]], align 8
-// CHECK17-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 6
-// CHECK17-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP5]], i64* [[TMP122]], align 8
-// CHECK17-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 6
-// CHECK17-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP5]], i64* [[TMP124]], align 8
-// CHECK17-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 6
-// CHECK17-NEXT:    store i8* null, i8** [[TMP125]], align 8
-// CHECK17-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 7
-// CHECK17-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to double**
-// CHECK17-NEXT:    store double* [[VLA1]], double** [[TMP127]], align 8
-// CHECK17-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 7
-// CHECK17-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to double**
-// CHECK17-NEXT:    store double* [[VLA1]], double** [[TMP129]], align 8
-// CHECK17-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK17-NEXT:    store i64 [[TMP88]], i64* [[TMP130]], align 8
-// CHECK17-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 7
+// CHECK17-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to float**
+// CHECK17-NEXT:    store float* [[VLA]], float** [[TMP122]], align 8
+// CHECK17-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to float**
+// CHECK17-NEXT:    store float* [[VLA]], float** [[TMP124]], align 8
+// CHECK17-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK17-NEXT:    store i64 [[TMP102]], i64* [[TMP125]], align 8
+// CHECK17-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK17-NEXT:    store i8* null, i8** [[TMP126]], align 8
+// CHECK17-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 4
+// CHECK17-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to [5 x [10 x double]]**
+// CHECK17-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP128]], align 8
+// CHECK17-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 4
+// CHECK17-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to [5 x [10 x double]]**
+// CHECK17-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP130]], align 8
+// CHECK17-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 4
 // CHECK17-NEXT:    store i8* null, i8** [[TMP131]], align 8
-// CHECK17-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 8
-// CHECK17-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to %struct.TT**
-// CHECK17-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP133]], align 8
-// CHECK17-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 8
-// CHECK17-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to %struct.TT**
-// CHECK17-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP135]], align 8
-// CHECK17-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 8
+// CHECK17-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 5
+// CHECK17-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i64*
+// CHECK17-NEXT:    store i64 5, i64* [[TMP133]], align 8
+// CHECK17-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 5
+// CHECK17-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i64*
+// CHECK17-NEXT:    store i64 5, i64* [[TMP135]], align 8
+// CHECK17-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 5
 // CHECK17-NEXT:    store i8* null, i8** [[TMP136]], align 8
-// CHECK17-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 9
+// CHECK17-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 6
 // CHECK17-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP84]], i64* [[TMP138]], align 8
-// CHECK17-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 9
+// CHECK17-NEXT:    store i64 [[TMP5]], i64* [[TMP138]], align 8
+// CHECK17-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 6
 // CHECK17-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP84]], i64* [[TMP140]], align 8
-// CHECK17-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 9
+// CHECK17-NEXT:    store i64 [[TMP5]], i64* [[TMP140]], align 8
+// CHECK17-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 6
 // CHECK17-NEXT:    store i8* null, i8** [[TMP141]], align 8
-// CHECK17-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP145:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170.region_id, i32 10, i8** [[TMP142]], i8** [[TMP143]], i64* [[TMP144]], i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK17-NEXT:    [[TMP146:%.*]] = icmp ne i32 [[TMP145]], 0
-// CHECK17-NEXT:    br i1 [[TMP146]], label [[OMP_OFFLOAD_FAILED23:%.*]], label [[OMP_OFFLOAD_CONT24:%.*]]
-// CHECK17:       omp_offload.failed23:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i64 [[TMP82]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP84]]) #[[ATTR4]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT24]]
-// CHECK17:       omp_offload.cont24:
-// CHECK17-NEXT:    br label [[OMP_IF_END26:%.*]]
-// CHECK17:       omp_if.else25:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i64 [[TMP82]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP84]]) #[[ATTR4]]
-// CHECK17-NEXT:    br label [[OMP_IF_END26]]
-// CHECK17:       omp_if.end26:
-// CHECK17-NEXT:    [[TMP147:%.*]] = load i32, i32* [[A]], align 4
-// CHECK17-NEXT:    [[TMP148:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP148]])
-// CHECK17-NEXT:    ret i32 [[TMP147]]
+// CHECK17-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 7
+// CHECK17-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to double**
+// CHECK17-NEXT:    store double* [[VLA1]], double** [[TMP143]], align 8
+// CHECK17-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 7
+// CHECK17-NEXT:    [[TMP145:%.*]] = bitcast i8** [[TMP144]] to double**
+// CHECK17-NEXT:    store double* [[VLA1]], double** [[TMP145]], align 8
+// CHECK17-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK17-NEXT:    store i64 [[TMP104]], i64* [[TMP146]], align 8
+// CHECK17-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 7
+// CHECK17-NEXT:    store i8* null, i8** [[TMP147]], align 8
+// CHECK17-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 8
+// CHECK17-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to %struct.TT**
+// CHECK17-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP149]], align 8
+// CHECK17-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 8
+// CHECK17-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to %struct.TT**
+// CHECK17-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP151]], align 8
+// CHECK17-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 8
+// CHECK17-NEXT:    store i8* null, i8** [[TMP152]], align 8
+// CHECK17-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 9
+// CHECK17-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP100]], i64* [[TMP154]], align 8
+// CHECK17-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 9
+// CHECK17-NEXT:    [[TMP156:%.*]] = bitcast i8** [[TMP155]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP100]], i64* [[TMP156]], align 8
+// CHECK17-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 9
+// CHECK17-NEXT:    store i8* null, i8** [[TMP157]], align 8
+// CHECK17-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK17-NEXT:    [[KERNEL_ARGS24:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP161]], align 4
+// CHECK17-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 10, i32* [[TMP162]], align 4
+// CHECK17-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP158]], i8*** [[TMP163]], align 8
+// CHECK17-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP159]], i8*** [[TMP164]], align 8
+// CHECK17-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP160]], i64** [[TMP165]], align 8
+// CHECK17-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP166]], align 8
+// CHECK17-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP167]], align 8
+// CHECK17-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP168]], align 8
+// CHECK17-NEXT:    [[TMP169:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]])
+// CHECK17-NEXT:    [[TMP170:%.*]] = icmp ne i32 [[TMP169]], 0
+// CHECK17-NEXT:    br i1 [[TMP170]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
+// CHECK17:       omp_offload.failed25:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i64 [[TMP98]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP100]]) #[[ATTR4]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
+// CHECK17:       omp_offload.cont26:
+// CHECK17-NEXT:    br label [[OMP_IF_END28:%.*]]
+// CHECK17:       omp_if.else27:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i64 [[TMP98]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP100]]) #[[ATTR4]]
+// CHECK17-NEXT:    br label [[OMP_IF_END28]]
+// CHECK17:       omp_if.end28:
+// CHECK17-NEXT:    [[TMP171:%.*]] = load i32, i32* [[A]], align 4
+// CHECK17-NEXT:    [[TMP172:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP172]])
+// CHECK17-NEXT:    ret i32 [[TMP171]]
 //
 //
 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103
@@ -5779,6 +6068,7 @@ int bar(int n){
 // CHECK17-NEXT:    [[AA_CASTED_I:%.*]] = alloca i64, align 8
 // CHECK17-NEXT:    [[LIN_CASTED_I:%.*]] = alloca i64, align 8
 // CHECK17-NEXT:    [[A_CASTED_I:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK17-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
 // CHECK17-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -5817,23 +6107,39 @@ int bar(int n){
 // CHECK17-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[TMP19]], i64 0, i64 0
 // CHECK17-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP12]], i32 0, i32 1
 // CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK17-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK17-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__3_EXIT:%.*]]
+// CHECK17-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP25]], align 4, !noalias !24
+// CHECK17-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP26]], align 4, !noalias !24
+// CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP20]], i8*** [[TMP27]], align 8, !noalias !24
+// CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP21]], i8*** [[TMP28]], align 8, !noalias !24
+// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP22]], i64** [[TMP29]], align 8, !noalias !24
+// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP30]], align 8, !noalias !24
+// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP31]], align 8, !noalias !24
+// CHECK17-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP32]], align 8, !noalias !24
+// CHECK17-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK17-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
+// CHECK17-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__3_EXIT:%.*]]
 // CHECK17:       omp_offload.failed.i:
-// CHECK17-NEXT:    [[TMP27:%.*]] = load i16, i16* [[TMP16]], align 2
+// CHECK17-NEXT:    [[TMP35:%.*]] = load i16, i16* [[TMP16]], align 2
 // CHECK17-NEXT:    [[CONV_I:%.*]] = bitcast i64* [[AA_CASTED_I]] to i16*
-// CHECK17-NEXT:    store i16 [[TMP27]], i16* [[CONV_I]], align 2, !noalias !24
-// CHECK17-NEXT:    [[TMP28:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !24
-// CHECK17-NEXT:    [[TMP29:%.*]] = load i32, i32* [[TMP23]], align 4
+// CHECK17-NEXT:    store i16 [[TMP35]], i16* [[CONV_I]], align 2, !noalias !24
+// CHECK17-NEXT:    [[TMP36:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !24
+// CHECK17-NEXT:    [[TMP37:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK17-NEXT:    [[CONV4_I:%.*]] = bitcast i64* [[LIN_CASTED_I]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP29]], i32* [[CONV4_I]], align 4, !noalias !24
-// CHECK17-NEXT:    [[TMP30:%.*]] = load i64, i64* [[LIN_CASTED_I]], align 8, !noalias !24
-// CHECK17-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP24]], align 4
+// CHECK17-NEXT:    store i32 [[TMP37]], i32* [[CONV4_I]], align 4, !noalias !24
+// CHECK17-NEXT:    [[TMP38:%.*]] = load i64, i64* [[LIN_CASTED_I]], align 8, !noalias !24
+// CHECK17-NEXT:    [[TMP39:%.*]] = load i32, i32* [[TMP24]], align 4
 // CHECK17-NEXT:    [[CONV5_I:%.*]] = bitcast i64* [[A_CASTED_I]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP31]], i32* [[CONV5_I]], align 4, !noalias !24
-// CHECK17-NEXT:    [[TMP32:%.*]] = load i64, i64* [[A_CASTED_I]], align 8, !noalias !24
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138(i64 [[TMP28]], i64 [[TMP30]], i64 [[TMP32]]) #[[ATTR4]]
+// CHECK17-NEXT:    store i32 [[TMP39]], i32* [[CONV5_I]], align 4, !noalias !24
+// CHECK17-NEXT:    [[TMP40:%.*]] = load i64, i64* [[A_CASTED_I]], align 8, !noalias !24
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138(i64 [[TMP36]], i64 [[TMP38]], i64 [[TMP40]]) #[[ATTR4]]
 // CHECK17-NEXT:    br label [[DOTOMP_OUTLINED__3_EXIT]]
 // CHECK17:       .omp_outlined..3.exit:
 // CHECK17-NEXT:    ret i32 0
@@ -6245,9 +6551,26 @@ int bar(int n){
 // CHECK17-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK17-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK17-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK17-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 8
+// CHECK17-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 8
+// CHECK17-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 8
+// CHECK17-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP46]], align 8
+// CHECK17-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK17-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK17-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -6257,15 +6580,15 @@ int bar(int n){
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK17-NEXT:    br label [[OMP_IF_END]]
 // CHECK17:       omp_if.end:
-// CHECK17-NEXT:    [[TMP42:%.*]] = mul nsw i64 1, [[TMP2]]
-// CHECK17-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP42]]
+// CHECK17-NEXT:    [[TMP50:%.*]] = mul nsw i64 1, [[TMP2]]
+// CHECK17-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP50]]
 // CHECK17-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
-// CHECK17-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK17-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK17-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK17-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP44]]
-// CHECK17-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK17-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK17-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK17-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK17-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP52]]
+// CHECK17-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK17-NEXT:    ret i32 [[ADD4]]
 //
 //
@@ -6337,9 +6660,26 @@ int bar(int n){
 // CHECK17-NEXT:    store i8* null, i8** [[TMP26]], align 8
 // CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK17-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK17-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK17-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK17-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK17-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK17-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK17-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK17-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -6349,8 +6689,8 @@ int bar(int n){
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK17-NEXT:    br label [[OMP_IF_END]]
 // CHECK17:       omp_if.end:
-// CHECK17-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
-// CHECK17-NEXT:    ret i32 [[TMP31]]
+// CHECK17-NEXT:    [[TMP39:%.*]] = load i32, i32* [[A]], align 4
+// CHECK17-NEXT:    ret i32 [[TMP39]]
 //
 //
 // CHECK17-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -6406,9 +6746,26 @@ int bar(int n){
 // CHECK17-NEXT:    store i8* null, i8** [[TMP19]], align 8
 // CHECK17-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK17-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK17-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK17-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK17-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK17-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK17-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK17-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -6418,8 +6775,8 @@ int bar(int n){
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK17-NEXT:    br label [[OMP_IF_END]]
 // CHECK17:       omp_if.end:
-// CHECK17-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK17-NEXT:    ret i32 [[TMP24]]
+// CHECK17-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK17-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242
@@ -6744,11 +7101,11 @@ int bar(int n){
 // CHECK19-NEXT:    [[DOTOFFLOAD_PTRS7:%.*]] = alloca [2 x i8*], align 4
 // CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS8:%.*]] = alloca [2 x i8*], align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[A_CASTED11:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[A_CASTED12:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS14:%.*]] = alloca [10 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS15:%.*]] = alloca [10 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS16:%.*]] = alloca [10 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [10 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [10 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [10 x i8*], align 4
 // CHECK19-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [10 x i64], align 4
 // CHECK19-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK19-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
@@ -6763,252 +7120,303 @@ int bar(int n){
 // CHECK19-NEXT:    [[TMP4:%.*]] = mul nuw i32 5, [[TMP3]]
 // CHECK19-NEXT:    [[VLA1:%.*]] = alloca double, i32 [[TMP4]], align 8
 // CHECK19-NEXT:    store i32 [[TMP3]], i32* [[__VLA_EXPR1]], align 4
-// CHECK19-NEXT:    [[TMP5:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0)
-// CHECK19-NEXT:    [[TMP6:%.*]] = icmp ne i32 [[TMP5]], 0
-// CHECK19-NEXT:    br i1 [[TMP6]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP5]], align 4
+// CHECK19-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 0, i32* [[TMP6]], align 4
+// CHECK19-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK19-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* null, i64** [[TMP9]], align 4
+// CHECK19-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* null, i64** [[TMP10]], align 4
+// CHECK19-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP11]], align 4
+// CHECK19-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP12]], align 4
+// CHECK19-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
+// CHECK19-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103() #[[ATTR4:[0-9]+]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
 // CHECK19-NEXT:    [[CALL:%.*]] = call noundef i64 @_Z7get_valv()
 // CHECK19-NEXT:    store i64 [[CALL]], i64* [[K]], align 8
-// CHECK19-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4
-// CHECK19-NEXT:    store i32 [[TMP7]], i32* [[A_CASTED]], align 4
-// CHECK19-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A_CASTED]], align 4
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110(i32 [[TMP8]], i64* [[K]]) #[[ATTR4]]
+// CHECK19-NEXT:    [[TMP15:%.*]] = load i32, i32* [[A]], align 4
+// CHECK19-NEXT:    store i32 [[TMP15]], i32* [[A_CASTED]], align 4
+// CHECK19-NEXT:    [[TMP16:%.*]] = load i32, i32* [[A_CASTED]], align 4
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l110(i32 [[TMP16]], i64* [[K]]) #[[ATTR4]]
 // CHECK19-NEXT:    store i32 12, i32* [[LIN]], align 4
-// CHECK19-NEXT:    [[TMP9:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK19-NEXT:    [[TMP17:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK19-NEXT:    [[CONV:%.*]] = bitcast i32* [[AA_CASTED]] to i16*
-// CHECK19-NEXT:    store i16 [[TMP9]], i16* [[CONV]], align 2
-// CHECK19-NEXT:    [[TMP10:%.*]] = load i32, i32* [[AA_CASTED]], align 4
-// CHECK19-NEXT:    [[TMP11:%.*]] = load i32, i32* [[LIN]], align 4
-// CHECK19-NEXT:    store i32 [[TMP11]], i32* [[LIN_CASTED]], align 4
-// CHECK19-NEXT:    [[TMP12:%.*]] = load i32, i32* [[LIN_CASTED]], align 4
-// CHECK19-NEXT:    [[TMP13:%.*]] = load i32, i32* [[A]], align 4
-// CHECK19-NEXT:    store i32 [[TMP13]], i32* [[A_CASTED2]], align 4
-// CHECK19-NEXT:    [[TMP14:%.*]] = load i32, i32* [[A_CASTED2]], align 4
-// CHECK19-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP10]], i32* [[TMP16]], align 4
-// CHECK19-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP10]], i32* [[TMP18]], align 4
-// CHECK19-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP19]], align 4
-// CHECK19-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP12]], i32* [[TMP21]], align 4
-// CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP12]], i32* [[TMP23]], align 4
-// CHECK19-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP24]], align 4
-// CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK19-NEXT:    store i16 [[TMP17]], i16* [[CONV]], align 2
+// CHECK19-NEXT:    [[TMP18:%.*]] = load i32, i32* [[AA_CASTED]], align 4
+// CHECK19-NEXT:    [[TMP19:%.*]] = load i32, i32* [[LIN]], align 4
+// CHECK19-NEXT:    store i32 [[TMP19]], i32* [[LIN_CASTED]], align 4
+// CHECK19-NEXT:    [[TMP20:%.*]] = load i32, i32* [[LIN_CASTED]], align 4
+// CHECK19-NEXT:    [[TMP21:%.*]] = load i32, i32* [[A]], align 4
+// CHECK19-NEXT:    store i32 [[TMP21]], i32* [[A_CASTED2]], align 4
+// CHECK19-NEXT:    [[TMP22:%.*]] = load i32, i32* [[A_CASTED2]], align 4
+// CHECK19-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP18]], i32* [[TMP24]], align 4
+// CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP14]], i32* [[TMP26]], align 4
-// CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP14]], i32* [[TMP28]], align 4
-// CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP29]], align 4
-// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP33:%.*]] = load i16, i16* [[AA]], align 2
-// CHECK19-NEXT:    store i16 [[TMP33]], i16* [[TMP32]], align 4
-// CHECK19-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP35:%.*]] = load i32, i32* [[LIN]], align 4
-// CHECK19-NEXT:    store i32 [[TMP35]], i32* [[TMP34]], align 4
-// CHECK19-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP37:%.*]] = load i32, i32* [[A]], align 4
-// CHECK19-NEXT:    store i32 [[TMP37]], i32* [[TMP36]], align 4
-// CHECK19-NEXT:    [[TMP38:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i32 1, i32 72, i32 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
-// CHECK19-NEXT:    [[TMP39:%.*]] = bitcast i8* [[TMP38]] to %struct.kmp_task_t_with_privates*
-// CHECK19-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP39]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP40]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP42:%.*]] = load i8*, i8** [[TMP41]], align 4
-// CHECK19-NEXT:    [[TMP43:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP42]], i8* align 4 [[TMP43]], i32 12, i1 false)
-// CHECK19-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP39]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP45:%.*]] = bitcast i8* [[TMP42]] to %struct.anon*
-// CHECK19-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP47:%.*]] = bitcast [3 x i64]* [[TMP46]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP47]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes to i8*), i32 24, i1 false)
-// CHECK19-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP49:%.*]] = bitcast [3 x i8*]* [[TMP48]] to i8*
-// CHECK19-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP30]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP49]], i8* align 4 [[TMP50]], i32 12, i1 false)
-// CHECK19-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP52:%.*]] = bitcast [3 x i8*]* [[TMP51]] to i8*
-// CHECK19-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP31]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP52]], i8* align 4 [[TMP53]], i32 12, i1 false)
-// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP44]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP55:%.*]] = load i16, i16* [[AA]], align 2
-// CHECK19-NEXT:    store i16 [[TMP55]], i16* [[TMP54]], align 4
-// CHECK19-NEXT:    [[TMP56:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i8* [[TMP38]])
-// CHECK19-NEXT:    [[TMP57:%.*]] = load i32, i32* [[A]], align 4
-// CHECK19-NEXT:    store i32 [[TMP57]], i32* [[A_CASTED3]], align 4
-// CHECK19-NEXT:    [[TMP58:%.*]] = load i32, i32* [[A_CASTED3]], align 4
-// CHECK19-NEXT:    [[TMP59:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK19-NEXT:    store i32 [[TMP18]], i32* [[TMP26]], align 4
+// CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP27]], align 4
+// CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP20]], i32* [[TMP29]], align 4
+// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP20]], i32* [[TMP31]], align 4
+// CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP32]], align 4
+// CHECK19-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP34:%.*]] = bitcast i8** [[TMP33]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP22]], i32* [[TMP34]], align 4
+// CHECK19-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP22]], i32* [[TMP36]], align 4
+// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP37]], align 4
+// CHECK19-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP41:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK19-NEXT:    store i16 [[TMP41]], i16* [[TMP40]], align 4
+// CHECK19-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP43:%.*]] = load i32, i32* [[LIN]], align 4
+// CHECK19-NEXT:    store i32 [[TMP43]], i32* [[TMP42]], align 4
+// CHECK19-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP45:%.*]] = load i32, i32* [[A]], align 4
+// CHECK19-NEXT:    store i32 [[TMP45]], i32* [[TMP44]], align 4
+// CHECK19-NEXT:    [[TMP46:%.*]] = call i8* @__kmpc_omp_target_task_alloc(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i32 1, i32 72, i32 12, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*), i64 -1)
+// CHECK19-NEXT:    [[TMP47:%.*]] = bitcast i8* [[TMP46]] to %struct.kmp_task_t_with_privates*
+// CHECK19-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP47]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP48]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP50:%.*]] = load i8*, i8** [[TMP49]], align 4
+// CHECK19-NEXT:    [[TMP51:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP50]], i8* align 4 [[TMP51]], i32 12, i1 false)
+// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP47]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP53:%.*]] = bitcast i8* [[TMP50]] to %struct.anon*
+// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP52]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP55:%.*]] = bitcast [3 x i64]* [[TMP54]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP55]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes to i8*), i32 24, i1 false)
+// CHECK19-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP52]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP57:%.*]] = bitcast [3 x i8*]* [[TMP56]] to i8*
+// CHECK19-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP38]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP57]], i8* align 4 [[TMP58]], i32 12, i1 false)
+// CHECK19-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP52]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP60:%.*]] = bitcast [3 x i8*]* [[TMP59]] to i8*
+// CHECK19-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP39]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP60]], i8* align 4 [[TMP61]], i32 12, i1 false)
+// CHECK19-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP52]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP63:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK19-NEXT:    store i16 [[TMP63]], i16* [[TMP62]], align 4
+// CHECK19-NEXT:    [[TMP64:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB2]], i32 [[TMP0]], i8* [[TMP46]])
+// CHECK19-NEXT:    [[TMP65:%.*]] = load i32, i32* [[A]], align 4
+// CHECK19-NEXT:    store i32 [[TMP65]], i32* [[A_CASTED3]], align 4
+// CHECK19-NEXT:    [[TMP66:%.*]] = load i32, i32* [[A_CASTED3]], align 4
+// CHECK19-NEXT:    [[TMP67:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK19-NEXT:    [[CONV5:%.*]] = bitcast i32* [[AA_CASTED4]] to i16*
-// CHECK19-NEXT:    store i16 [[TMP59]], i16* [[CONV5]], align 2
-// CHECK19-NEXT:    [[TMP60:%.*]] = load i32, i32* [[AA_CASTED4]], align 4
-// CHECK19-NEXT:    [[TMP61:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK19-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP61]], 10
+// CHECK19-NEXT:    store i16 [[TMP67]], i16* [[CONV5]], align 2
+// CHECK19-NEXT:    [[TMP68:%.*]] = load i32, i32* [[AA_CASTED4]], align 4
+// CHECK19-NEXT:    [[TMP69:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK19-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP69]], 10
 // CHECK19-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK19:       omp_if.then:
-// CHECK19-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP63:%.*]] = bitcast i8** [[TMP62]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP58]], i32* [[TMP63]], align 4
-// CHECK19-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP65:%.*]] = bitcast i8** [[TMP64]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP58]], i32* [[TMP65]], align 4
-// CHECK19-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP66]], align 4
-// CHECK19-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP60]], i32* [[TMP68]], align 4
-// CHECK19-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP60]], i32* [[TMP70]], align 4
-// CHECK19-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP71]], align 4
-// CHECK19-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146.region_id, i32 2, i8** [[TMP72]], i8** [[TMP73]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK19-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
-// CHECK19-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED9:%.*]], label [[OMP_OFFLOAD_CONT10:%.*]]
-// CHECK19:       omp_offload.failed9:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i32 [[TMP58]], i32 [[TMP60]]) #[[ATTR4]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT10]]
-// CHECK19:       omp_offload.cont10:
+// CHECK19-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP66]], i32* [[TMP71]], align 4
+// CHECK19-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP66]], i32* [[TMP73]], align 4
+// CHECK19-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP74]], align 4
+// CHECK19-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP68]], i32* [[TMP76]], align 4
+// CHECK19-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP68]], i32* [[TMP78]], align 4
+// CHECK19-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP79]], align 4
+// CHECK19-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
+// CHECK19-NEXT:    [[KERNEL_ARGS9:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP82]], align 4
+// CHECK19-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 2, i32* [[TMP83]], align 4
+// CHECK19-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP80]], i8*** [[TMP84]], align 4
+// CHECK19-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP81]], i8*** [[TMP85]], align 4
+// CHECK19-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP86]], align 4
+// CHECK19-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP87]], align 4
+// CHECK19-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP88]], align 4
+// CHECK19-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP89]], align 4
+// CHECK19-NEXT:    [[TMP90:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]])
+// CHECK19-NEXT:    [[TMP91:%.*]] = icmp ne i32 [[TMP90]], 0
+// CHECK19-NEXT:    br i1 [[TMP91]], label [[OMP_OFFLOAD_FAILED10:%.*]], label [[OMP_OFFLOAD_CONT11:%.*]]
+// CHECK19:       omp_offload.failed10:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i32 [[TMP66]], i32 [[TMP68]]) #[[ATTR4]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT11]]
+// CHECK19:       omp_offload.cont11:
 // CHECK19-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK19:       omp_if.else:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i32 [[TMP58]], i32 [[TMP60]]) #[[ATTR4]]
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l146(i32 [[TMP66]], i32 [[TMP68]]) #[[ATTR4]]
 // CHECK19-NEXT:    br label [[OMP_IF_END]]
 // CHECK19:       omp_if.end:
-// CHECK19-NEXT:    [[TMP76:%.*]] = load i32, i32* [[A]], align 4
-// CHECK19-NEXT:    store i32 [[TMP76]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK19-NEXT:    [[TMP77:%.*]] = load i32, i32* [[A]], align 4
-// CHECK19-NEXT:    store i32 [[TMP77]], i32* [[A_CASTED11]], align 4
-// CHECK19-NEXT:    [[TMP78:%.*]] = load i32, i32* [[A_CASTED11]], align 4
-// CHECK19-NEXT:    [[TMP79:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK19-NEXT:    store i32 [[TMP79]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK19-NEXT:    [[TMP80:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK19-NEXT:    [[TMP81:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK19-NEXT:    [[CMP12:%.*]] = icmp sgt i32 [[TMP81]], 20
-// CHECK19-NEXT:    br i1 [[CMP12]], label [[OMP_IF_THEN13:%.*]], label [[OMP_IF_ELSE19:%.*]]
-// CHECK19:       omp_if.then13:
-// CHECK19-NEXT:    [[TMP82:%.*]] = mul nuw i32 [[TMP1]], 4
-// CHECK19-NEXT:    [[TMP83:%.*]] = sext i32 [[TMP82]] to i64
-// CHECK19-NEXT:    [[TMP84:%.*]] = mul nuw i32 5, [[TMP3]]
-// CHECK19-NEXT:    [[TMP85:%.*]] = mul nuw i32 [[TMP84]], 8
-// CHECK19-NEXT:    [[TMP86:%.*]] = sext i32 [[TMP85]] to i64
-// CHECK19-NEXT:    [[TMP87:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP87]], i8* align 4 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i32 80, i1 false)
-// CHECK19-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP78]], i32* [[TMP89]], align 4
-// CHECK19-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP78]], i32* [[TMP91]], align 4
-// CHECK19-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP92]], align 4
-// CHECK19-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to [10 x float]**
-// CHECK19-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP94]], align 4
-// CHECK19-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to [10 x float]**
-// CHECK19-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP96]], align 4
-// CHECK19-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP97]], align 4
-// CHECK19-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP1]], i32* [[TMP99]], align 4
-// CHECK19-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP1]], i32* [[TMP101]], align 4
-// CHECK19-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP102]], align 4
-// CHECK19-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to float**
-// CHECK19-NEXT:    store float* [[VLA]], float** [[TMP104]], align 4
-// CHECK19-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP106:%.*]] = bitcast i8** [[TMP105]] to float**
-// CHECK19-NEXT:    store float* [[VLA]], float** [[TMP106]], align 4
-// CHECK19-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK19-NEXT:    store i64 [[TMP83]], i64* [[TMP107]], align 4
-// CHECK19-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP92:%.*]] = load i32, i32* [[A]], align 4
+// CHECK19-NEXT:    store i32 [[TMP92]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK19-NEXT:    [[TMP93:%.*]] = load i32, i32* [[A]], align 4
+// CHECK19-NEXT:    store i32 [[TMP93]], i32* [[A_CASTED12]], align 4
+// CHECK19-NEXT:    [[TMP94:%.*]] = load i32, i32* [[A_CASTED12]], align 4
+// CHECK19-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK19-NEXT:    store i32 [[TMP95]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK19-NEXT:    [[TMP96:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK19-NEXT:    [[TMP97:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK19-NEXT:    [[CMP13:%.*]] = icmp sgt i32 [[TMP97]], 20
+// CHECK19-NEXT:    br i1 [[CMP13]], label [[OMP_IF_THEN14:%.*]], label [[OMP_IF_ELSE21:%.*]]
+// CHECK19:       omp_if.then14:
+// CHECK19-NEXT:    [[TMP98:%.*]] = mul nuw i32 [[TMP1]], 4
+// CHECK19-NEXT:    [[TMP99:%.*]] = sext i32 [[TMP98]] to i64
+// CHECK19-NEXT:    [[TMP100:%.*]] = mul nuw i32 5, [[TMP3]]
+// CHECK19-NEXT:    [[TMP101:%.*]] = mul nuw i32 [[TMP100]], 8
+// CHECK19-NEXT:    [[TMP102:%.*]] = sext i32 [[TMP101]] to i64
+// CHECK19-NEXT:    [[TMP103:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP103]], i8* align 4 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i32 80, i1 false)
+// CHECK19-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP94]], i32* [[TMP105]], align 4
+// CHECK19-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP94]], i32* [[TMP107]], align 4
+// CHECK19-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 0
 // CHECK19-NEXT:    store i8* null, i8** [[TMP108]], align 4
-// CHECK19-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 4
-// CHECK19-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to [5 x [10 x double]]**
-// CHECK19-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP110]], align 4
-// CHECK19-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 4
-// CHECK19-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to [5 x [10 x double]]**
-// CHECK19-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP112]], align 4
-// CHECK19-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 4
+// CHECK19-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to [10 x float]**
+// CHECK19-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP110]], align 4
+// CHECK19-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to [10 x float]**
+// CHECK19-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP112]], align 4
+// CHECK19-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 1
 // CHECK19-NEXT:    store i8* null, i8** [[TMP113]], align 4
-// CHECK19-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 5
+// CHECK19-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 2
 // CHECK19-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32*
-// CHECK19-NEXT:    store i32 5, i32* [[TMP115]], align 4
-// CHECK19-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 5
+// CHECK19-NEXT:    store i32 [[TMP1]], i32* [[TMP115]], align 4
+// CHECK19-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 2
 // CHECK19-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i32*
-// CHECK19-NEXT:    store i32 5, i32* [[TMP117]], align 4
-// CHECK19-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 5
+// CHECK19-NEXT:    store i32 [[TMP1]], i32* [[TMP117]], align 4
+// CHECK19-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 2
 // CHECK19-NEXT:    store i8* null, i8** [[TMP118]], align 4
-// CHECK19-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 6
-// CHECK19-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP3]], i32* [[TMP120]], align 4
-// CHECK19-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 6
-// CHECK19-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP3]], i32* [[TMP122]], align 4
-// CHECK19-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 6
-// CHECK19-NEXT:    store i8* null, i8** [[TMP123]], align 4
-// CHECK19-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 7
-// CHECK19-NEXT:    [[TMP125:%.*]] = bitcast i8** [[TMP124]] to double**
-// CHECK19-NEXT:    store double* [[VLA1]], double** [[TMP125]], align 4
-// CHECK19-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 7
-// CHECK19-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to double**
-// CHECK19-NEXT:    store double* [[VLA1]], double** [[TMP127]], align 4
-// CHECK19-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK19-NEXT:    store i64 [[TMP86]], i64* [[TMP128]], align 4
-// CHECK19-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 7
+// CHECK19-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to float**
+// CHECK19-NEXT:    store float* [[VLA]], float** [[TMP120]], align 4
+// CHECK19-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to float**
+// CHECK19-NEXT:    store float* [[VLA]], float** [[TMP122]], align 4
+// CHECK19-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK19-NEXT:    store i64 [[TMP99]], i64* [[TMP123]], align 4
+// CHECK19-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 3
+// CHECK19-NEXT:    store i8* null, i8** [[TMP124]], align 4
+// CHECK19-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 4
+// CHECK19-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to [5 x [10 x double]]**
+// CHECK19-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP126]], align 4
+// CHECK19-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 4
+// CHECK19-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to [5 x [10 x double]]**
+// CHECK19-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP128]], align 4
+// CHECK19-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 4
 // CHECK19-NEXT:    store i8* null, i8** [[TMP129]], align 4
-// CHECK19-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 8
-// CHECK19-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to %struct.TT**
-// CHECK19-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP131]], align 4
-// CHECK19-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 8
-// CHECK19-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to %struct.TT**
-// CHECK19-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP133]], align 4
-// CHECK19-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 8
+// CHECK19-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 5
+// CHECK19-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i32*
+// CHECK19-NEXT:    store i32 5, i32* [[TMP131]], align 4
+// CHECK19-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 5
+// CHECK19-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i32*
+// CHECK19-NEXT:    store i32 5, i32* [[TMP133]], align 4
+// CHECK19-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 5
 // CHECK19-NEXT:    store i8* null, i8** [[TMP134]], align 4
-// CHECK19-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 9
+// CHECK19-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 6
 // CHECK19-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP80]], i32* [[TMP136]], align 4
-// CHECK19-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 9
+// CHECK19-NEXT:    store i32 [[TMP3]], i32* [[TMP136]], align 4
+// CHECK19-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 6
 // CHECK19-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP80]], i32* [[TMP138]], align 4
-// CHECK19-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 9
+// CHECK19-NEXT:    store i32 [[TMP3]], i32* [[TMP138]], align 4
+// CHECK19-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 6
 // CHECK19-NEXT:    store i8* null, i8** [[TMP139]], align 4
-// CHECK19-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP143:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170.region_id, i32 10, i8** [[TMP140]], i8** [[TMP141]], i64* [[TMP142]], i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK19-NEXT:    [[TMP144:%.*]] = icmp ne i32 [[TMP143]], 0
-// CHECK19-NEXT:    br i1 [[TMP144]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK19:       omp_offload.failed17:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i32 [[TMP78]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP80]]) #[[ATTR4]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK19:       omp_offload.cont18:
-// CHECK19-NEXT:    br label [[OMP_IF_END20:%.*]]
-// CHECK19:       omp_if.else19:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i32 [[TMP78]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP80]]) #[[ATTR4]]
-// CHECK19-NEXT:    br label [[OMP_IF_END20]]
-// CHECK19:       omp_if.end20:
-// CHECK19-NEXT:    [[TMP145:%.*]] = load i32, i32* [[A]], align 4
-// CHECK19-NEXT:    [[TMP146:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP146]])
-// CHECK19-NEXT:    ret i32 [[TMP145]]
+// CHECK19-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 7
+// CHECK19-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to double**
+// CHECK19-NEXT:    store double* [[VLA1]], double** [[TMP141]], align 4
+// CHECK19-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 7
+// CHECK19-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to double**
+// CHECK19-NEXT:    store double* [[VLA1]], double** [[TMP143]], align 4
+// CHECK19-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK19-NEXT:    store i64 [[TMP102]], i64* [[TMP144]], align 4
+// CHECK19-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 7
+// CHECK19-NEXT:    store i8* null, i8** [[TMP145]], align 4
+// CHECK19-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 8
+// CHECK19-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to %struct.TT**
+// CHECK19-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP147]], align 4
+// CHECK19-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 8
+// CHECK19-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to %struct.TT**
+// CHECK19-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP149]], align 4
+// CHECK19-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 8
+// CHECK19-NEXT:    store i8* null, i8** [[TMP150]], align 4
+// CHECK19-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 9
+// CHECK19-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP96]], i32* [[TMP152]], align 4
+// CHECK19-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 9
+// CHECK19-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP96]], i32* [[TMP154]], align 4
+// CHECK19-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 9
+// CHECK19-NEXT:    store i8* null, i8** [[TMP155]], align 4
+// CHECK19-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK19-NEXT:    [[KERNEL_ARGS18:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP159]], align 4
+// CHECK19-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 10, i32* [[TMP160]], align 4
+// CHECK19-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP156]], i8*** [[TMP161]], align 4
+// CHECK19-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP157]], i8*** [[TMP162]], align 4
+// CHECK19-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP158]], i64** [[TMP163]], align 4
+// CHECK19-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP164]], align 4
+// CHECK19-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP165]], align 4
+// CHECK19-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP166]], align 4
+// CHECK19-NEXT:    [[TMP167:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]])
+// CHECK19-NEXT:    [[TMP168:%.*]] = icmp ne i32 [[TMP167]], 0
+// CHECK19-NEXT:    br i1 [[TMP168]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
+// CHECK19:       omp_offload.failed19:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i32 [[TMP94]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP96]]) #[[ATTR4]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
+// CHECK19:       omp_offload.cont20:
+// CHECK19-NEXT:    br label [[OMP_IF_END22:%.*]]
+// CHECK19:       omp_if.else21:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l170(i32 [[TMP94]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP96]]) #[[ATTR4]]
+// CHECK19-NEXT:    br label [[OMP_IF_END22]]
+// CHECK19:       omp_if.end22:
+// CHECK19-NEXT:    [[TMP169:%.*]] = load i32, i32* [[A]], align 4
+// CHECK19-NEXT:    [[TMP170:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP170]])
+// CHECK19-NEXT:    ret i32 [[TMP169]]
 //
 //
 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103
@@ -7376,6 +7784,7 @@ int bar(int n){
 // CHECK19-NEXT:    [[AA_CASTED_I:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[LIN_CASTED_I:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[A_CASTED_I:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK19-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 4
 // CHECK19-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -7414,21 +7823,37 @@ int bar(int n){
 // CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[TMP19]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP12]], i32 0, i32 1
 // CHECK19-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK19-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK19-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__3_EXIT:%.*]]
+// CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP25]], align 4, !noalias !25
+// CHECK19-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP26]], align 4, !noalias !25
+// CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP20]], i8*** [[TMP27]], align 4, !noalias !25
+// CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP21]], i8*** [[TMP28]], align 4, !noalias !25
+// CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP22]], i64** [[TMP29]], align 4, !noalias !25
+// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP30]], align 4, !noalias !25
+// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP31]], align 4, !noalias !25
+// CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP32]], align 4, !noalias !25
+// CHECK19-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK19-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
+// CHECK19-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__3_EXIT:%.*]]
 // CHECK19:       omp_offload.failed.i:
-// CHECK19-NEXT:    [[TMP27:%.*]] = load i16, i16* [[TMP16]], align 2
+// CHECK19-NEXT:    [[TMP35:%.*]] = load i16, i16* [[TMP16]], align 2
 // CHECK19-NEXT:    [[CONV_I:%.*]] = bitcast i32* [[AA_CASTED_I]] to i16*
-// CHECK19-NEXT:    store i16 [[TMP27]], i16* [[CONV_I]], align 2, !noalias !25
-// CHECK19-NEXT:    [[TMP28:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !25
-// CHECK19-NEXT:    [[TMP29:%.*]] = load i32, i32* [[TMP23]], align 4
-// CHECK19-NEXT:    store i32 [[TMP29]], i32* [[LIN_CASTED_I]], align 4, !noalias !25
-// CHECK19-NEXT:    [[TMP30:%.*]] = load i32, i32* [[LIN_CASTED_I]], align 4, !noalias !25
-// CHECK19-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP24]], align 4
-// CHECK19-NEXT:    store i32 [[TMP31]], i32* [[A_CASTED_I]], align 4, !noalias !25
-// CHECK19-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A_CASTED_I]], align 4, !noalias !25
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138(i32 [[TMP28]], i32 [[TMP30]], i32 [[TMP32]]) #[[ATTR4]]
+// CHECK19-NEXT:    store i16 [[TMP35]], i16* [[CONV_I]], align 2, !noalias !25
+// CHECK19-NEXT:    [[TMP36:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !25
+// CHECK19-NEXT:    [[TMP37:%.*]] = load i32, i32* [[TMP23]], align 4
+// CHECK19-NEXT:    store i32 [[TMP37]], i32* [[LIN_CASTED_I]], align 4, !noalias !25
+// CHECK19-NEXT:    [[TMP38:%.*]] = load i32, i32* [[LIN_CASTED_I]], align 4, !noalias !25
+// CHECK19-NEXT:    [[TMP39:%.*]] = load i32, i32* [[TMP24]], align 4
+// CHECK19-NEXT:    store i32 [[TMP39]], i32* [[A_CASTED_I]], align 4, !noalias !25
+// CHECK19-NEXT:    [[TMP40:%.*]] = load i32, i32* [[A_CASTED_I]], align 4, !noalias !25
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l138(i32 [[TMP36]], i32 [[TMP38]], i32 [[TMP40]]) #[[ATTR4]]
 // CHECK19-NEXT:    br label [[DOTOMP_OUTLINED__3_EXIT]]
 // CHECK19:       .omp_outlined..3.exit:
 // CHECK19-NEXT:    ret i32 0
@@ -7830,9 +8255,26 @@ int bar(int n){
 // CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK19-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK19-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK19-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 4
+// CHECK19-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 4
+// CHECK19-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 4
+// CHECK19-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK19-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP46]], align 4
+// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK19-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK19-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -7842,15 +8284,15 @@ int bar(int n){
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK19-NEXT:    br label [[OMP_IF_END]]
 // CHECK19:       omp_if.end:
-// CHECK19-NEXT:    [[TMP42:%.*]] = mul nsw i32 1, [[TMP1]]
-// CHECK19-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP42]]
+// CHECK19-NEXT:    [[TMP50:%.*]] = mul nsw i32 1, [[TMP1]]
+// CHECK19-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP50]]
 // CHECK19-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
-// CHECK19-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK19-NEXT:    [[CONV:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK19-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK19-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP44]]
-// CHECK19-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK19-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK19-NEXT:    [[CONV:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK19-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK19-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP52]]
+// CHECK19-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK19-NEXT:    ret i32 [[ADD3]]
 //
 //
@@ -7921,9 +8363,26 @@ int bar(int n){
 // CHECK19-NEXT:    store i8* null, i8** [[TMP26]], align 4
 // CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK19-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK19-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK19-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK19-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK19-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK19-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK19-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK19-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -7933,8 +8392,8 @@ int bar(int n){
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l224(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK19-NEXT:    br label [[OMP_IF_END]]
 // CHECK19:       omp_if.end:
-// CHECK19-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
-// CHECK19-NEXT:    ret i32 [[TMP31]]
+// CHECK19-NEXT:    [[TMP39:%.*]] = load i32, i32* [[A]], align 4
+// CHECK19-NEXT:    ret i32 [[TMP39]]
 //
 //
 // CHECK19-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -7989,9 +8448,26 @@ int bar(int n){
 // CHECK19-NEXT:    store i8* null, i8** [[TMP19]], align 4
 // CHECK19-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK19-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK19-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK19-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK19-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 4
+// CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 4
+// CHECK19-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK19-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK19-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -8001,8 +8477,8 @@ int bar(int n){
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l207(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK19-NEXT:    br label [[OMP_IF_END]]
 // CHECK19:       omp_if.end:
-// CHECK19-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK19-NEXT:    ret i32 [[TMP24]]
+// CHECK19-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK19-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l242

diff --git a/clang/test/OpenMP/target_parallel_for_depend_codegen.cpp b/clang/test/OpenMP/target_parallel_for_depend_codegen.cpp
index 95cab30..193d2c5 100644 (file)
--- a/clang/test/OpenMP/target_parallel_for_depend_codegen.cpp
+++ b/clang/test/OpenMP/target_parallel_for_depend_codegen.cpp
@@ -182,7 +182,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 0
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // CHECK:       [[FAIL]]
@@ -200,7 +200,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 2
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 2, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SZT]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 0, i32 0, i8* null, i32 0, i8* null)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 
 // CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]

diff --git a/clang/test/OpenMP/target_parallel_for_simd_codegen.cpp b/clang/test/OpenMP/target_parallel_for_simd_codegen.cpp
index cb2193f..902e692 100644 (file)
--- a/clang/test/OpenMP/target_parallel_for_simd_codegen.cpp
+++ b/clang/test/OpenMP/target_parallel_for_simd_codegen.cpp
@@ -328,11 +328,11 @@ int bar(int n){
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [2 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [2 x i8*], align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[A_CASTED15:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[A_CASTED16:%.*]] = alloca i64, align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [10 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [10 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [10 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [10 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [10 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [10 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [10 x i64], align 8
 // CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
@@ -402,177 +402,228 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP35]], align 8
 // CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l108.region_id, i32 3, i8** [[TMP36]], i8** [[TMP37]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
-// CHECK1-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP38]], align 4
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP39]], align 4
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP36]], i8*** [[TMP40]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP37]], i8*** [[TMP41]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP47:%.*]] = icmp ne i32 [[TMP46]], 0
+// CHECK1-NEXT:    br i1 [[TMP47]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l108(i64 [[TMP16]], i64 [[TMP18]], i64 [[TMP20]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP40:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP48:%.*]] = load i32, i32* [[A]], align 4
 // CHECK1-NEXT:    [[CONV7:%.*]] = bitcast i64* [[A_CASTED6]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP40]], i32* [[CONV7]], align 4
-// CHECK1-NEXT:    [[TMP41:%.*]] = load i64, i64* [[A_CASTED6]], align 8
-// CHECK1-NEXT:    [[TMP42:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK1-NEXT:    store i32 [[TMP48]], i32* [[CONV7]], align 4
+// CHECK1-NEXT:    [[TMP49:%.*]] = load i64, i64* [[A_CASTED6]], align 8
+// CHECK1-NEXT:    [[TMP50:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK1-NEXT:    [[CONV9:%.*]] = bitcast i64* [[AA_CASTED8]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP42]], i16* [[CONV9]], align 2
-// CHECK1-NEXT:    [[TMP43:%.*]] = load i64, i64* [[AA_CASTED8]], align 8
-// CHECK1-NEXT:    [[TMP44:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP44]], 10
+// CHECK1-NEXT:    store i16 [[TMP50]], i16* [[CONV9]], align 2
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i64, i64* [[AA_CASTED8]], align 8
+// CHECK1-NEXT:    [[TMP52:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP52]], 10
 // CHECK1-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK1:       omp_if.then:
-// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP41]], i64* [[TMP46]], align 8
-// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP41]], i64* [[TMP48]], align 8
-// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP49]], align 8
-// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP43]], i64* [[TMP51]], align 8
-// CHECK1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP43]], i64* [[TMP53]], align 8
-// CHECK1-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP54]], align 8
-// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
-// CHECK1-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
-// CHECK1:       omp_offload.failed13:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i64 [[TMP41]], i64 [[TMP43]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
-// CHECK1:       omp_offload.cont14:
+// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP49]], i64* [[TMP54]], align 8
+// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP49]], i64* [[TMP56]], align 8
+// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP57]], align 8
+// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP51]], i64* [[TMP59]], align 8
+// CHECK1-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP51]], i64* [[TMP61]], align 8
+// CHECK1-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP62]], align 8
+// CHECK1-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS13:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP65]], align 4
+// CHECK1-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP66]], align 4
+// CHECK1-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP63]], i8*** [[TMP67]], align 8
+// CHECK1-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 8
+// CHECK1-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP69]], align 8
+// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP70]], align 8
+// CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP71]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP72]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]])
+// CHECK1-NEXT:    [[TMP74:%.*]] = icmp ne i32 [[TMP73]], 0
+// CHECK1-NEXT:    br i1 [[TMP74]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
+// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i64 [[TMP49]], i64 [[TMP51]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
+// CHECK1:       omp_offload.cont15:
 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK1:       omp_if.else:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i64 [[TMP41]], i64 [[TMP43]]) #[[ATTR4]]
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i64 [[TMP49]], i64 [[TMP51]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP59:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    store i32 [[TMP59]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK1-NEXT:    [[TMP60:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    [[CONV16:%.*]] = bitcast i64* [[A_CASTED15]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP60]], i32* [[CONV16]], align 4
-// CHECK1-NEXT:    [[TMP61:%.*]] = load i64, i64* [[A_CASTED15]], align 8
-// CHECK1-NEXT:    [[TMP62:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK1-NEXT:    [[CONV17:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP62]], i32* [[CONV17]], align 4
-// CHECK1-NEXT:    [[TMP63:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK1-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP18:%.*]] = icmp sgt i32 [[TMP64]], 20
-// CHECK1-NEXT:    br i1 [[CMP18]], label [[OMP_IF_THEN19:%.*]], label [[OMP_IF_ELSE25:%.*]]
-// CHECK1:       omp_if.then19:
-// CHECK1-NEXT:    [[TMP65:%.*]] = mul nuw i64 [[TMP2]], 4
-// CHECK1-NEXT:    [[TMP66:%.*]] = mul nuw i64 5, [[TMP5]]
-// CHECK1-NEXT:    [[TMP67:%.*]] = mul nuw i64 [[TMP66]], 8
-// CHECK1-NEXT:    [[TMP68:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP68]], i8* align 8 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i64 80, i1 false)
-// CHECK1-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP61]], i64* [[TMP70]], align 8
-// CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP61]], i64* [[TMP72]], align 8
-// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP73]], align 8
-// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to [10 x float]**
-// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP75]], align 8
-// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to [10 x float]**
-// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP77]], align 8
-// CHECK1-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP78]], align 8
-// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP80]], align 8
-// CHECK1-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP82]], align 8
-// CHECK1-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP83]], align 8
-// CHECK1-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to float**
-// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP85]], align 8
-// CHECK1-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to float**
-// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP87]], align 8
-// CHECK1-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK1-NEXT:    store i64 [[TMP65]], i64* [[TMP88]], align 8
-// CHECK1-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 3
+// CHECK1-NEXT:    [[TMP75:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    store i32 [[TMP75]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT:    [[TMP76:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[CONV17:%.*]] = bitcast i64* [[A_CASTED16]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP76]], i32* [[CONV17]], align 4
+// CHECK1-NEXT:    [[TMP77:%.*]] = load i64, i64* [[A_CASTED16]], align 8
+// CHECK1-NEXT:    [[TMP78:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT:    [[CONV18:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP78]], i32* [[CONV18]], align 4
+// CHECK1-NEXT:    [[TMP79:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK1-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP19:%.*]] = icmp sgt i32 [[TMP80]], 20
+// CHECK1-NEXT:    br i1 [[CMP19]], label [[OMP_IF_THEN20:%.*]], label [[OMP_IF_ELSE27:%.*]]
+// CHECK1:       omp_if.then20:
+// CHECK1-NEXT:    [[TMP81:%.*]] = mul nuw i64 [[TMP2]], 4
+// CHECK1-NEXT:    [[TMP82:%.*]] = mul nuw i64 5, [[TMP5]]
+// CHECK1-NEXT:    [[TMP83:%.*]] = mul nuw i64 [[TMP82]], 8
+// CHECK1-NEXT:    [[TMP84:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP84]], i8* align 8 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i64 80, i1 false)
+// CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP77]], i64* [[TMP86]], align 8
+// CHECK1-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP88:%.*]] = bitcast i8** [[TMP87]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP77]], i64* [[TMP88]], align 8
+// CHECK1-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
 // CHECK1-NEXT:    store i8* null, i8** [[TMP89]], align 8
-// CHECK1-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to [5 x [10 x double]]**
-// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP91]], align 8
-// CHECK1-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to [5 x [10 x double]]**
-// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP93]], align 8
-// CHECK1-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 4
+// CHECK1-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to [10 x float]**
+// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP91]], align 8
+// CHECK1-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to [10 x float]**
+// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP93]], align 8
+// CHECK1-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
 // CHECK1-NEXT:    store i8* null, i8** [[TMP94]], align 8
-// CHECK1-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i64*
-// CHECK1-NEXT:    store i64 5, i64* [[TMP96]], align 8
-// CHECK1-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 5
+// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP96]], align 8
+// CHECK1-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i64*
-// CHECK1-NEXT:    store i64 5, i64* [[TMP98]], align 8
-// CHECK1-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 5
+// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP98]], align 8
+// CHECK1-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
 // CHECK1-NEXT:    store i8* null, i8** [[TMP99]], align 8
-// CHECK1-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP101]], align 8
-// CHECK1-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP103]], align 8
-// CHECK1-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 6
-// CHECK1-NEXT:    store i8* null, i8** [[TMP104]], align 8
-// CHECK1-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP106:%.*]] = bitcast i8** [[TMP105]] to double**
-// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP106]], align 8
-// CHECK1-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to double**
-// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP108]], align 8
-// CHECK1-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK1-NEXT:    store i64 [[TMP67]], i64* [[TMP109]], align 8
-// CHECK1-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 7
+// CHECK1-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to float**
+// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP101]], align 8
+// CHECK1-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to float**
+// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP103]], align 8
+// CHECK1-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK1-NEXT:    store i64 [[TMP81]], i64* [[TMP104]], align 8
+// CHECK1-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK1-NEXT:    store i8* null, i8** [[TMP105]], align 8
+// CHECK1-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to [5 x [10 x double]]**
+// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP107]], align 8
+// CHECK1-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to [5 x [10 x double]]**
+// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP109]], align 8
+// CHECK1-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 4
 // CHECK1-NEXT:    store i8* null, i8** [[TMP110]], align 8
-// CHECK1-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to %struct.TT**
-// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP112]], align 8
-// CHECK1-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to %struct.TT**
-// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP114]], align 8
-// CHECK1-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 8
+// CHECK1-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i64*
+// CHECK1-NEXT:    store i64 5, i64* [[TMP112]], align 8
+// CHECK1-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i64*
+// CHECK1-NEXT:    store i64 5, i64* [[TMP114]], align 8
+// CHECK1-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 5
 // CHECK1-NEXT:    store i8* null, i8** [[TMP115]], align 8
-// CHECK1-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 9
+// CHECK1-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 6
 // CHECK1-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP63]], i64* [[TMP117]], align 8
-// CHECK1-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 9
+// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP117]], align 8
+// CHECK1-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 6
 // CHECK1-NEXT:    [[TMP119:%.*]] = bitcast i8** [[TMP118]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP63]], i64* [[TMP119]], align 8
-// CHECK1-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 9
+// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP119]], align 8
+// CHECK1-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 6
 // CHECK1-NEXT:    store i8* null, i8** [[TMP120]], align 8
-// CHECK1-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP124:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140.region_id, i32 10, i8** [[TMP121]], i8** [[TMP122]], i64* [[TMP123]], i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP125:%.*]] = icmp ne i32 [[TMP124]], 0
-// CHECK1-NEXT:    br i1 [[TMP125]], label [[OMP_OFFLOAD_FAILED23:%.*]], label [[OMP_OFFLOAD_CONT24:%.*]]
-// CHECK1:       omp_offload.failed23:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i64 [[TMP61]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP63]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT24]]
-// CHECK1:       omp_offload.cont24:
-// CHECK1-NEXT:    br label [[OMP_IF_END26:%.*]]
-// CHECK1:       omp_if.else25:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i64 [[TMP61]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP63]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_IF_END26]]
-// CHECK1:       omp_if.end26:
-// CHECK1-NEXT:    [[TMP126:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    [[TMP127:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP127]])
-// CHECK1-NEXT:    ret i32 [[TMP126]]
+// CHECK1-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to double**
+// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP122]], align 8
+// CHECK1-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to double**
+// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP124]], align 8
+// CHECK1-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK1-NEXT:    store i64 [[TMP83]], i64* [[TMP125]], align 8
+// CHECK1-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 7
+// CHECK1-NEXT:    store i8* null, i8** [[TMP126]], align 8
+// CHECK1-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to %struct.TT**
+// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP128]], align 8
+// CHECK1-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to %struct.TT**
+// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP130]], align 8
+// CHECK1-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 8
+// CHECK1-NEXT:    store i8* null, i8** [[TMP131]], align 8
+// CHECK1-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 9
+// CHECK1-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP79]], i64* [[TMP133]], align 8
+// CHECK1-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 9
+// CHECK1-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP79]], i64* [[TMP135]], align 8
+// CHECK1-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 9
+// CHECK1-NEXT:    store i8* null, i8** [[TMP136]], align 8
+// CHECK1-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS24:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP140]], align 4
+// CHECK1-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 10, i32* [[TMP141]], align 4
+// CHECK1-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP137]], i8*** [[TMP142]], align 8
+// CHECK1-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP138]], i8*** [[TMP143]], align 8
+// CHECK1-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP139]], i64** [[TMP144]], align 8
+// CHECK1-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP145]], align 8
+// CHECK1-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP146]], align 8
+// CHECK1-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP147]], align 8
+// CHECK1-NEXT:    [[TMP148:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]])
+// CHECK1-NEXT:    [[TMP149:%.*]] = icmp ne i32 [[TMP148]], 0
+// CHECK1-NEXT:    br i1 [[TMP149]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
+// CHECK1:       omp_offload.failed25:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i64 [[TMP77]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP79]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
+// CHECK1:       omp_offload.cont26:
+// CHECK1-NEXT:    br label [[OMP_IF_END28:%.*]]
+// CHECK1:       omp_if.else27:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i64 [[TMP77]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP79]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_IF_END28]]
+// CHECK1:       omp_if.end28:
+// CHECK1-NEXT:    [[TMP150:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP151:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP151]])
+// CHECK1-NEXT:    ret i32 [[TMP150]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96
@@ -658,6 +709,7 @@ int bar(int n){
 // CHECK1-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
 // CHECK1-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
 // CHECK1-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 8
+// CHECK1-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK1-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -681,9 +733,25 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !25
 // CHECK1-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !25
 // CHECK1-NEXT:    [[TMP10:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !25
-// CHECK1-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK1-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK1-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP11]], align 4, !noalias !25
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP12]], align 4, !noalias !25
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8, !noalias !25
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8, !noalias !25
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8, !noalias !25
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP16]], align 8, !noalias !25
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8, !noalias !25
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP18]], align 8, !noalias !25
+// CHECK1-NEXT:    [[TMP19:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK1-NEXT:    [[TMP20:%.*]] = icmp ne i32 [[TMP19]], 0
+// CHECK1-NEXT:    br i1 [[TMP20]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK1:       omp_offload.failed.i:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96() #[[ATTR4]]
 // CHECK1-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
@@ -1372,9 +1440,26 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK1-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK1-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1384,15 +1469,15 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP42:%.*]] = mul nsw i64 1, [[TMP2]]
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP42]]
+// CHECK1-NEXT:    [[TMP50:%.*]] = mul nsw i64 1, [[TMP2]]
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP50]]
 // CHECK1-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
-// CHECK1-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK1-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP44]]
-// CHECK1-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK1-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP52]]
+// CHECK1-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK1-NEXT:    ret i32 [[ADD4]]
 //
 //
@@ -1464,9 +1549,26 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
 // CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK1-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK1-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1476,8 +1578,8 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP31]]
+// CHECK1-NEXT:    [[TMP39:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP39]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -1533,9 +1635,26 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP19]], align 8
 // CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK1-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK1-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1545,8 +1664,8 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP24]]
+// CHECK1-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216
@@ -1885,11 +2004,11 @@ int bar(int n){
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS7:%.*]] = alloca [2 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS8:%.*]] = alloca [2 x i8*], align 4
 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[A_CASTED11:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[A_CASTED12:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS14:%.*]] = alloca [10 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS15:%.*]] = alloca [10 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS16:%.*]] = alloca [10 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [10 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [10 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [10 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [10 x i64], align 4
 // CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
@@ -1951,176 +2070,227 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP31]], align 4
 // CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l108.region_id, i32 3, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK3-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP32]], i8*** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP33]], i8*** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 4
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK3-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l108(i32 [[TMP12]], i32 [[TMP14]], i32 [[TMP16]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP36:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP36]], i32* [[A_CASTED3]], align 4
-// CHECK3-NEXT:    [[TMP37:%.*]] = load i32, i32* [[A_CASTED3]], align 4
-// CHECK3-NEXT:    [[TMP38:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK3-NEXT:    [[TMP44:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP44]], i32* [[A_CASTED3]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = load i32, i32* [[A_CASTED3]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK3-NEXT:    [[CONV5:%.*]] = bitcast i32* [[AA_CASTED4]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP38]], i16* [[CONV5]], align 2
-// CHECK3-NEXT:    [[TMP39:%.*]] = load i32, i32* [[AA_CASTED4]], align 4
-// CHECK3-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP40]], 10
+// CHECK3-NEXT:    store i16 [[TMP46]], i16* [[CONV5]], align 2
+// CHECK3-NEXT:    [[TMP47:%.*]] = load i32, i32* [[AA_CASTED4]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP48]], 10
 // CHECK3-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK3:       omp_if.then:
-// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP37]], i32* [[TMP42]], align 4
-// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP37]], i32* [[TMP44]], align 4
-// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP45]], align 4
-// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP39]], i32* [[TMP47]], align 4
-// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP39]], i32* [[TMP49]], align 4
-// CHECK3-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP50]], align 4
-// CHECK3-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP53:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116.region_id, i32 2, i8** [[TMP51]], i8** [[TMP52]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP54:%.*]] = icmp ne i32 [[TMP53]], 0
-// CHECK3-NEXT:    br i1 [[TMP54]], label [[OMP_OFFLOAD_FAILED9:%.*]], label [[OMP_OFFLOAD_CONT10:%.*]]
-// CHECK3:       omp_offload.failed9:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i32 [[TMP37]], i32 [[TMP39]]) #[[ATTR4]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT10]]
-// CHECK3:       omp_offload.cont10:
+// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP45]], i32* [[TMP50]], align 4
+// CHECK3-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP45]], i32* [[TMP52]], align 4
+// CHECK3-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP53]], align 4
+// CHECK3-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP47]], i32* [[TMP55]], align 4
+// CHECK3-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP47]], i32* [[TMP57]], align 4
+// CHECK3-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP58]], align 4
+// CHECK3-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS9:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP61]], align 4
+// CHECK3-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP62]], align 4
+// CHECK3-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP59]], i8*** [[TMP63]], align 4
+// CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP60]], i8*** [[TMP64]], align 4
+// CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP65]], align 4
+// CHECK3-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP66]], align 4
+// CHECK3-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP67]], align 4
+// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP68]], align 4
+// CHECK3-NEXT:    [[TMP69:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]])
+// CHECK3-NEXT:    [[TMP70:%.*]] = icmp ne i32 [[TMP69]], 0
+// CHECK3-NEXT:    br i1 [[TMP70]], label [[OMP_OFFLOAD_FAILED10:%.*]], label [[OMP_OFFLOAD_CONT11:%.*]]
+// CHECK3:       omp_offload.failed10:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i32 [[TMP45]], i32 [[TMP47]]) #[[ATTR4]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT11]]
+// CHECK3:       omp_offload.cont11:
 // CHECK3-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK3:       omp_if.else:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i32 [[TMP37]], i32 [[TMP39]]) #[[ATTR4]]
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i32 [[TMP45]], i32 [[TMP47]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP55:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK3-NEXT:    [[TMP56:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP56]], i32* [[A_CASTED11]], align 4
-// CHECK3-NEXT:    [[TMP57:%.*]] = load i32, i32* [[A_CASTED11]], align 4
-// CHECK3-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK3-NEXT:    store i32 [[TMP58]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK3-NEXT:    [[TMP59:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK3-NEXT:    [[TMP60:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP12:%.*]] = icmp sgt i32 [[TMP60]], 20
-// CHECK3-NEXT:    br i1 [[CMP12]], label [[OMP_IF_THEN13:%.*]], label [[OMP_IF_ELSE19:%.*]]
-// CHECK3:       omp_if.then13:
-// CHECK3-NEXT:    [[TMP61:%.*]] = mul nuw i32 [[TMP1]], 4
-// CHECK3-NEXT:    [[TMP62:%.*]] = sext i32 [[TMP61]] to i64
-// CHECK3-NEXT:    [[TMP63:%.*]] = mul nuw i32 5, [[TMP3]]
-// CHECK3-NEXT:    [[TMP64:%.*]] = mul nuw i32 [[TMP63]], 8
-// CHECK3-NEXT:    [[TMP65:%.*]] = sext i32 [[TMP64]] to i64
-// CHECK3-NEXT:    [[TMP66:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP66]], i8* align 4 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i32 80, i1 false)
-// CHECK3-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP57]], i32* [[TMP68]], align 4
-// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP57]], i32* [[TMP70]], align 4
-// CHECK3-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP71]], align 4
-// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to [10 x float]**
-// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP73]], align 4
-// CHECK3-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to [10 x float]**
-// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP75]], align 4
-// CHECK3-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP76]], align 4
-// CHECK3-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP78]], align 4
-// CHECK3-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP80]], align 4
-// CHECK3-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP81]], align 4
-// CHECK3-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to float**
-// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP83]], align 4
-// CHECK3-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to float**
-// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP85]], align 4
-// CHECK3-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK3-NEXT:    store i64 [[TMP62]], i64* [[TMP86]], align 4
-// CHECK3-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP71:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP71]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK3-NEXT:    [[TMP72:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP72]], i32* [[A_CASTED12]], align 4
+// CHECK3-NEXT:    [[TMP73:%.*]] = load i32, i32* [[A_CASTED12]], align 4
+// CHECK3-NEXT:    [[TMP74:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK3-NEXT:    store i32 [[TMP74]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK3-NEXT:    [[TMP75:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK3-NEXT:    [[TMP76:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP13:%.*]] = icmp sgt i32 [[TMP76]], 20
+// CHECK3-NEXT:    br i1 [[CMP13]], label [[OMP_IF_THEN14:%.*]], label [[OMP_IF_ELSE21:%.*]]
+// CHECK3:       omp_if.then14:
+// CHECK3-NEXT:    [[TMP77:%.*]] = mul nuw i32 [[TMP1]], 4
+// CHECK3-NEXT:    [[TMP78:%.*]] = sext i32 [[TMP77]] to i64
+// CHECK3-NEXT:    [[TMP79:%.*]] = mul nuw i32 5, [[TMP3]]
+// CHECK3-NEXT:    [[TMP80:%.*]] = mul nuw i32 [[TMP79]], 8
+// CHECK3-NEXT:    [[TMP81:%.*]] = sext i32 [[TMP80]] to i64
+// CHECK3-NEXT:    [[TMP82:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP82]], i8* align 4 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i32 80, i1 false)
+// CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP73]], i32* [[TMP84]], align 4
+// CHECK3-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP73]], i32* [[TMP86]], align 4
+// CHECK3-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 0
 // CHECK3-NEXT:    store i8* null, i8** [[TMP87]], align 4
-// CHECK3-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to [5 x [10 x double]]**
-// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP89]], align 4
-// CHECK3-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to [5 x [10 x double]]**
-// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP91]], align 4
-// CHECK3-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to [10 x float]**
+// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP89]], align 4
+// CHECK3-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to [10 x float]**
+// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP91]], align 4
+// CHECK3-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 1
 // CHECK3-NEXT:    store i8* null, i8** [[TMP92]], align 4
-// CHECK3-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32*
-// CHECK3-NEXT:    store i32 5, i32* [[TMP94]], align 4
-// CHECK3-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP94]], align 4
+// CHECK3-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32*
-// CHECK3-NEXT:    store i32 5, i32* [[TMP96]], align 4
-// CHECK3-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 5
+// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP96]], align 4
+// CHECK3-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 2
 // CHECK3-NEXT:    store i8* null, i8** [[TMP97]], align 4
-// CHECK3-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 6
-// CHECK3-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP99]], align 4
-// CHECK3-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 6
-// CHECK3-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP101]], align 4
-// CHECK3-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 6
-// CHECK3-NEXT:    store i8* null, i8** [[TMP102]], align 4
-// CHECK3-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 7
-// CHECK3-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to double**
-// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP104]], align 4
-// CHECK3-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 7
-// CHECK3-NEXT:    [[TMP106:%.*]] = bitcast i8** [[TMP105]] to double**
-// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP106]], align 4
-// CHECK3-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK3-NEXT:    store i64 [[TMP65]], i64* [[TMP107]], align 4
-// CHECK3-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to float**
+// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP99]], align 4
+// CHECK3-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to float**
+// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP101]], align 4
+// CHECK3-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK3-NEXT:    store i64 [[TMP78]], i64* [[TMP102]], align 4
+// CHECK3-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 3
+// CHECK3-NEXT:    store i8* null, i8** [[TMP103]], align 4
+// CHECK3-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to [5 x [10 x double]]**
+// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP105]], align 4
+// CHECK3-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to [5 x [10 x double]]**
+// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP107]], align 4
+// CHECK3-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 4
 // CHECK3-NEXT:    store i8* null, i8** [[TMP108]], align 4
-// CHECK3-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 8
-// CHECK3-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to %struct.TT**
-// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP110]], align 4
-// CHECK3-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 8
-// CHECK3-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to %struct.TT**
-// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP112]], align 4
-// CHECK3-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32*
+// CHECK3-NEXT:    store i32 5, i32* [[TMP110]], align 4
+// CHECK3-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32*
+// CHECK3-NEXT:    store i32 5, i32* [[TMP112]], align 4
+// CHECK3-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 5
 // CHECK3-NEXT:    store i8* null, i8** [[TMP113]], align 4
-// CHECK3-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 9
+// CHECK3-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 6
 // CHECK3-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP59]], i32* [[TMP115]], align 4
-// CHECK3-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 9
+// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP115]], align 4
+// CHECK3-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 6
 // CHECK3-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP59]], i32* [[TMP117]], align 4
-// CHECK3-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 9
+// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP117]], align 4
+// CHECK3-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 6
 // CHECK3-NEXT:    store i8* null, i8** [[TMP118]], align 4
-// CHECK3-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP122:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140.region_id, i32 10, i8** [[TMP119]], i8** [[TMP120]], i64* [[TMP121]], i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP123:%.*]] = icmp ne i32 [[TMP122]], 0
-// CHECK3-NEXT:    br i1 [[TMP123]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK3:       omp_offload.failed17:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i32 [[TMP57]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP59]]) #[[ATTR4]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK3:       omp_offload.cont18:
-// CHECK3-NEXT:    br label [[OMP_IF_END20:%.*]]
-// CHECK3:       omp_if.else19:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i32 [[TMP57]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP59]]) #[[ATTR4]]
-// CHECK3-NEXT:    br label [[OMP_IF_END20]]
-// CHECK3:       omp_if.end20:
-// CHECK3-NEXT:    [[TMP124:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    [[TMP125:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP125]])
-// CHECK3-NEXT:    ret i32 [[TMP124]]
+// CHECK3-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to double**
+// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP120]], align 4
+// CHECK3-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to double**
+// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP122]], align 4
+// CHECK3-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK3-NEXT:    store i64 [[TMP81]], i64* [[TMP123]], align 4
+// CHECK3-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 7
+// CHECK3-NEXT:    store i8* null, i8** [[TMP124]], align 4
+// CHECK3-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to %struct.TT**
+// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP126]], align 4
+// CHECK3-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to %struct.TT**
+// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP128]], align 4
+// CHECK3-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 8
+// CHECK3-NEXT:    store i8* null, i8** [[TMP129]], align 4
+// CHECK3-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 9
+// CHECK3-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP75]], i32* [[TMP131]], align 4
+// CHECK3-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 9
+// CHECK3-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP75]], i32* [[TMP133]], align 4
+// CHECK3-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 9
+// CHECK3-NEXT:    store i8* null, i8** [[TMP134]], align 4
+// CHECK3-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS18:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP138]], align 4
+// CHECK3-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 10, i32* [[TMP139]], align 4
+// CHECK3-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP135]], i8*** [[TMP140]], align 4
+// CHECK3-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP136]], i8*** [[TMP141]], align 4
+// CHECK3-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP137]], i64** [[TMP142]], align 4
+// CHECK3-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP143]], align 4
+// CHECK3-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP144]], align 4
+// CHECK3-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP145]], align 4
+// CHECK3-NEXT:    [[TMP146:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]])
+// CHECK3-NEXT:    [[TMP147:%.*]] = icmp ne i32 [[TMP146]], 0
+// CHECK3-NEXT:    br i1 [[TMP147]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
+// CHECK3:       omp_offload.failed19:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i32 [[TMP73]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP75]]) #[[ATTR4]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
+// CHECK3:       omp_offload.cont20:
+// CHECK3-NEXT:    br label [[OMP_IF_END22:%.*]]
+// CHECK3:       omp_if.else21:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i32 [[TMP73]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP75]]) #[[ATTR4]]
+// CHECK3-NEXT:    br label [[OMP_IF_END22]]
+// CHECK3:       omp_if.end22:
+// CHECK3-NEXT:    [[TMP148:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    [[TMP149:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP149]])
+// CHECK3-NEXT:    ret i32 [[TMP148]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96
@@ -2206,6 +2376,7 @@ int bar(int n){
 // CHECK3-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 4
 // CHECK3-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 4
 // CHECK3-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 4
+// CHECK3-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK3-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 4
 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -2229,9 +2400,25 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 4, !noalias !26
 // CHECK3-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !26
 // CHECK3-NEXT:    [[TMP10:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !26
-// CHECK3-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK3-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK3-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP11]], align 4, !noalias !26
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP12]], align 4, !noalias !26
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4, !noalias !26
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4, !noalias !26
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP15]], align 4, !noalias !26
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP16]], align 4, !noalias !26
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 4, !noalias !26
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP18]], align 4, !noalias !26
+// CHECK3-NEXT:    [[TMP19:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK3-NEXT:    [[TMP20:%.*]] = icmp ne i32 [[TMP19]], 0
+// CHECK3-NEXT:    br i1 [[TMP20]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK3:       omp_offload.failed.i:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96() #[[ATTR4]]
 // CHECK3-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
@@ -2899,9 +3086,26 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK3-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK3-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2911,15 +3115,15 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP42:%.*]] = mul nsw i32 1, [[TMP1]]
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP42]]
+// CHECK3-NEXT:    [[TMP50:%.*]] = mul nsw i32 1, [[TMP1]]
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP50]]
 // CHECK3-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
-// CHECK3-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK3-NEXT:    [[CONV:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK3-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP44]]
-// CHECK3-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK3-NEXT:    [[CONV:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK3-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP52]]
+// CHECK3-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK3-NEXT:    ret i32 [[ADD3]]
 //
 //
@@ -2990,9 +3194,26 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP26]], align 4
 // CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK3-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK3-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3002,8 +3223,8 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP31]]
+// CHECK3-NEXT:    [[TMP39:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP39]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -3058,9 +3279,26 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP19]], align 4
 // CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK3-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK3-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3070,8 +3308,8 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP24]]
+// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l216
@@ -3402,11 +3640,11 @@ int bar(int n){
 // CHECK5-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [2 x i8*], align 8
 // CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [2 x i8*], align 8
 // CHECK5-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[A_CASTED15:%.*]] = alloca i64, align 8
+// CHECK5-NEXT:    [[A_CASTED16:%.*]] = alloca i64, align 8
 // CHECK5-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [10 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [10 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [10 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [10 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [10 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [10 x i8*], align 8
 // CHECK5-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [10 x i64], align 8
 // CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK5-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
@@ -3476,177 +3714,228 @@ int bar(int n){
 // CHECK5-NEXT:    store i8* null, i8** [[TMP35]], align 8
 // CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l108.region_id, i32 3, i8** [[TMP36]], i8** [[TMP37]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK5-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
-// CHECK5-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP38]], align 4
+// CHECK5-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 3, i32* [[TMP39]], align 4
+// CHECK5-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP36]], i8*** [[TMP40]], align 8
+// CHECK5-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP37]], i8*** [[TMP41]], align 8
+// CHECK5-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP42]], align 8
+// CHECK5-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP43]], align 8
+// CHECK5-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP44]], align 8
+// CHECK5-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP45]], align 8
+// CHECK5-NEXT:    [[TMP46:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP47:%.*]] = icmp ne i32 [[TMP46]], 0
+// CHECK5-NEXT:    br i1 [[TMP47]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l108(i64 [[TMP16]], i64 [[TMP18]], i64 [[TMP20]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
-// CHECK5-NEXT:    [[TMP40:%.*]] = load i32, i32* [[A]], align 4
+// CHECK5-NEXT:    [[TMP48:%.*]] = load i32, i32* [[A]], align 4
 // CHECK5-NEXT:    [[CONV7:%.*]] = bitcast i64* [[A_CASTED6]] to i32*
-// CHECK5-NEXT:    store i32 [[TMP40]], i32* [[CONV7]], align 4
-// CHECK5-NEXT:    [[TMP41:%.*]] = load i64, i64* [[A_CASTED6]], align 8
-// CHECK5-NEXT:    [[TMP42:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK5-NEXT:    store i32 [[TMP48]], i32* [[CONV7]], align 4
+// CHECK5-NEXT:    [[TMP49:%.*]] = load i64, i64* [[A_CASTED6]], align 8
+// CHECK5-NEXT:    [[TMP50:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK5-NEXT:    [[CONV9:%.*]] = bitcast i64* [[AA_CASTED8]] to i16*
-// CHECK5-NEXT:    store i16 [[TMP42]], i16* [[CONV9]], align 2
-// CHECK5-NEXT:    [[TMP43:%.*]] = load i64, i64* [[AA_CASTED8]], align 8
-// CHECK5-NEXT:    [[TMP44:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP44]], 10
+// CHECK5-NEXT:    store i16 [[TMP50]], i16* [[CONV9]], align 2
+// CHECK5-NEXT:    [[TMP51:%.*]] = load i64, i64* [[AA_CASTED8]], align 8
+// CHECK5-NEXT:    [[TMP52:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP52]], 10
 // CHECK5-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK5:       omp_if.then:
-// CHECK5-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP41]], i64* [[TMP46]], align 8
-// CHECK5-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP41]], i64* [[TMP48]], align 8
-// CHECK5-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP49]], align 8
-// CHECK5-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 1
-// CHECK5-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP43]], i64* [[TMP51]], align 8
-// CHECK5-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 1
-// CHECK5-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP43]], i64* [[TMP53]], align 8
-// CHECK5-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 1
-// CHECK5-NEXT:    store i8* null, i8** [[TMP54]], align 8
-// CHECK5-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK5-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
-// CHECK5-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
-// CHECK5:       omp_offload.failed13:
-// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i64 [[TMP41]], i64 [[TMP43]]) #[[ATTR4]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
-// CHECK5:       omp_offload.cont14:
+// CHECK5-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP49]], i64* [[TMP54]], align 8
+// CHECK5-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP49]], i64* [[TMP56]], align 8
+// CHECK5-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP57]], align 8
+// CHECK5-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 1
+// CHECK5-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP51]], i64* [[TMP59]], align 8
+// CHECK5-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 1
+// CHECK5-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP51]], i64* [[TMP61]], align 8
+// CHECK5-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 1
+// CHECK5-NEXT:    store i8* null, i8** [[TMP62]], align 8
+// CHECK5-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    [[KERNEL_ARGS13:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP65]], align 4
+// CHECK5-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 2, i32* [[TMP66]], align 4
+// CHECK5-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP63]], i8*** [[TMP67]], align 8
+// CHECK5-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 8
+// CHECK5-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP69]], align 8
+// CHECK5-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP70]], align 8
+// CHECK5-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP71]], align 8
+// CHECK5-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP72]], align 8
+// CHECK5-NEXT:    [[TMP73:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS13]])
+// CHECK5-NEXT:    [[TMP74:%.*]] = icmp ne i32 [[TMP73]], 0
+// CHECK5-NEXT:    br i1 [[TMP74]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
+// CHECK5:       omp_offload.failed14:
+// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i64 [[TMP49]], i64 [[TMP51]]) #[[ATTR4]]
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
+// CHECK5:       omp_offload.cont15:
 // CHECK5-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK5:       omp_if.else:
-// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i64 [[TMP41]], i64 [[TMP43]]) #[[ATTR4]]
+// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i64 [[TMP49]], i64 [[TMP51]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_IF_END]]
 // CHECK5:       omp_if.end:
-// CHECK5-NEXT:    [[TMP59:%.*]] = load i32, i32* [[A]], align 4
-// CHECK5-NEXT:    store i32 [[TMP59]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK5-NEXT:    [[TMP60:%.*]] = load i32, i32* [[A]], align 4
-// CHECK5-NEXT:    [[CONV16:%.*]] = bitcast i64* [[A_CASTED15]] to i32*
-// CHECK5-NEXT:    store i32 [[TMP60]], i32* [[CONV16]], align 4
-// CHECK5-NEXT:    [[TMP61:%.*]] = load i64, i64* [[A_CASTED15]], align 8
-// CHECK5-NEXT:    [[TMP62:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK5-NEXT:    [[CONV17:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK5-NEXT:    store i32 [[TMP62]], i32* [[CONV17]], align 4
-// CHECK5-NEXT:    [[TMP63:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK5-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK5-NEXT:    [[CMP18:%.*]] = icmp sgt i32 [[TMP64]], 20
-// CHECK5-NEXT:    br i1 [[CMP18]], label [[OMP_IF_THEN19:%.*]], label [[OMP_IF_ELSE25:%.*]]
-// CHECK5:       omp_if.then19:
-// CHECK5-NEXT:    [[TMP65:%.*]] = mul nuw i64 [[TMP2]], 4
-// CHECK5-NEXT:    [[TMP66:%.*]] = mul nuw i64 5, [[TMP5]]
-// CHECK5-NEXT:    [[TMP67:%.*]] = mul nuw i64 [[TMP66]], 8
-// CHECK5-NEXT:    [[TMP68:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK5-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP68]], i8* align 8 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i64 80, i1 false)
-// CHECK5-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP61]], i64* [[TMP70]], align 8
-// CHECK5-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP61]], i64* [[TMP72]], align 8
-// CHECK5-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP73]], align 8
-// CHECK5-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK5-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to [10 x float]**
-// CHECK5-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP75]], align 8
-// CHECK5-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK5-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to [10 x float]**
-// CHECK5-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP77]], align 8
-// CHECK5-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 1
-// CHECK5-NEXT:    store i8* null, i8** [[TMP78]], align 8
-// CHECK5-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
-// CHECK5-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP2]], i64* [[TMP80]], align 8
-// CHECK5-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
-// CHECK5-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP2]], i64* [[TMP82]], align 8
-// CHECK5-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 2
-// CHECK5-NEXT:    store i8* null, i8** [[TMP83]], align 8
-// CHECK5-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
-// CHECK5-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to float**
-// CHECK5-NEXT:    store float* [[VLA]], float** [[TMP85]], align 8
-// CHECK5-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
-// CHECK5-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to float**
-// CHECK5-NEXT:    store float* [[VLA]], float** [[TMP87]], align 8
-// CHECK5-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK5-NEXT:    store i64 [[TMP65]], i64* [[TMP88]], align 8
-// CHECK5-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 3
+// CHECK5-NEXT:    [[TMP75:%.*]] = load i32, i32* [[A]], align 4
+// CHECK5-NEXT:    store i32 [[TMP75]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK5-NEXT:    [[TMP76:%.*]] = load i32, i32* [[A]], align 4
+// CHECK5-NEXT:    [[CONV17:%.*]] = bitcast i64* [[A_CASTED16]] to i32*
+// CHECK5-NEXT:    store i32 [[TMP76]], i32* [[CONV17]], align 4
+// CHECK5-NEXT:    [[TMP77:%.*]] = load i64, i64* [[A_CASTED16]], align 8
+// CHECK5-NEXT:    [[TMP78:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK5-NEXT:    [[CONV18:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
+// CHECK5-NEXT:    store i32 [[TMP78]], i32* [[CONV18]], align 4
+// CHECK5-NEXT:    [[TMP79:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK5-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK5-NEXT:    [[CMP19:%.*]] = icmp sgt i32 [[TMP80]], 20
+// CHECK5-NEXT:    br i1 [[CMP19]], label [[OMP_IF_THEN20:%.*]], label [[OMP_IF_ELSE27:%.*]]
+// CHECK5:       omp_if.then20:
+// CHECK5-NEXT:    [[TMP81:%.*]] = mul nuw i64 [[TMP2]], 4
+// CHECK5-NEXT:    [[TMP82:%.*]] = mul nuw i64 5, [[TMP5]]
+// CHECK5-NEXT:    [[TMP83:%.*]] = mul nuw i64 [[TMP82]], 8
+// CHECK5-NEXT:    [[TMP84:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK5-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP84]], i8* align 8 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i64 80, i1 false)
+// CHECK5-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP77]], i64* [[TMP86]], align 8
+// CHECK5-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP88:%.*]] = bitcast i8** [[TMP87]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP77]], i64* [[TMP88]], align 8
+// CHECK5-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
 // CHECK5-NEXT:    store i8* null, i8** [[TMP89]], align 8
-// CHECK5-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 4
-// CHECK5-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to [5 x [10 x double]]**
-// CHECK5-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP91]], align 8
-// CHECK5-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 4
-// CHECK5-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to [5 x [10 x double]]**
-// CHECK5-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP93]], align 8
-// CHECK5-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 4
+// CHECK5-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
+// CHECK5-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to [10 x float]**
+// CHECK5-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP91]], align 8
+// CHECK5-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
+// CHECK5-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to [10 x float]**
+// CHECK5-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP93]], align 8
+// CHECK5-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
 // CHECK5-NEXT:    store i8* null, i8** [[TMP94]], align 8
-// CHECK5-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 5
+// CHECK5-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
 // CHECK5-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i64*
-// CHECK5-NEXT:    store i64 5, i64* [[TMP96]], align 8
-// CHECK5-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 5
+// CHECK5-NEXT:    store i64 [[TMP2]], i64* [[TMP96]], align 8
+// CHECK5-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
 // CHECK5-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i64*
-// CHECK5-NEXT:    store i64 5, i64* [[TMP98]], align 8
-// CHECK5-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 5
+// CHECK5-NEXT:    store i64 [[TMP2]], i64* [[TMP98]], align 8
+// CHECK5-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
 // CHECK5-NEXT:    store i8* null, i8** [[TMP99]], align 8
-// CHECK5-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 6
-// CHECK5-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP5]], i64* [[TMP101]], align 8
-// CHECK5-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 6
-// CHECK5-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP5]], i64* [[TMP103]], align 8
-// CHECK5-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 6
-// CHECK5-NEXT:    store i8* null, i8** [[TMP104]], align 8
-// CHECK5-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 7
-// CHECK5-NEXT:    [[TMP106:%.*]] = bitcast i8** [[TMP105]] to double**
-// CHECK5-NEXT:    store double* [[VLA1]], double** [[TMP106]], align 8
-// CHECK5-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 7
-// CHECK5-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to double**
-// CHECK5-NEXT:    store double* [[VLA1]], double** [[TMP108]], align 8
-// CHECK5-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK5-NEXT:    store i64 [[TMP67]], i64* [[TMP109]], align 8
-// CHECK5-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 7
+// CHECK5-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
+// CHECK5-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to float**
+// CHECK5-NEXT:    store float* [[VLA]], float** [[TMP101]], align 8
+// CHECK5-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
+// CHECK5-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to float**
+// CHECK5-NEXT:    store float* [[VLA]], float** [[TMP103]], align 8
+// CHECK5-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK5-NEXT:    store i64 [[TMP81]], i64* [[TMP104]], align 8
+// CHECK5-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK5-NEXT:    store i8* null, i8** [[TMP105]], align 8
+// CHECK5-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 4
+// CHECK5-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to [5 x [10 x double]]**
+// CHECK5-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP107]], align 8
+// CHECK5-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 4
+// CHECK5-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to [5 x [10 x double]]**
+// CHECK5-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP109]], align 8
+// CHECK5-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 4
 // CHECK5-NEXT:    store i8* null, i8** [[TMP110]], align 8
-// CHECK5-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 8
-// CHECK5-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to %struct.TT**
-// CHECK5-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP112]], align 8
-// CHECK5-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 8
-// CHECK5-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to %struct.TT**
-// CHECK5-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP114]], align 8
-// CHECK5-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 8
+// CHECK5-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 5
+// CHECK5-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i64*
+// CHECK5-NEXT:    store i64 5, i64* [[TMP112]], align 8
+// CHECK5-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 5
+// CHECK5-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i64*
+// CHECK5-NEXT:    store i64 5, i64* [[TMP114]], align 8
+// CHECK5-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 5
 // CHECK5-NEXT:    store i8* null, i8** [[TMP115]], align 8
-// CHECK5-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 9
+// CHECK5-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 6
 // CHECK5-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP63]], i64* [[TMP117]], align 8
-// CHECK5-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 9
+// CHECK5-NEXT:    store i64 [[TMP5]], i64* [[TMP117]], align 8
+// CHECK5-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 6
 // CHECK5-NEXT:    [[TMP119:%.*]] = bitcast i8** [[TMP118]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP63]], i64* [[TMP119]], align 8
-// CHECK5-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 9
+// CHECK5-NEXT:    store i64 [[TMP5]], i64* [[TMP119]], align 8
+// CHECK5-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 6
 // CHECK5-NEXT:    store i8* null, i8** [[TMP120]], align 8
-// CHECK5-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP124:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140.region_id, i32 10, i8** [[TMP121]], i8** [[TMP122]], i64* [[TMP123]], i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK5-NEXT:    [[TMP125:%.*]] = icmp ne i32 [[TMP124]], 0
-// CHECK5-NEXT:    br i1 [[TMP125]], label [[OMP_OFFLOAD_FAILED23:%.*]], label [[OMP_OFFLOAD_CONT24:%.*]]
-// CHECK5:       omp_offload.failed23:
-// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i64 [[TMP61]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP63]]) #[[ATTR4]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT24]]
-// CHECK5:       omp_offload.cont24:
-// CHECK5-NEXT:    br label [[OMP_IF_END26:%.*]]
-// CHECK5:       omp_if.else25:
-// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i64 [[TMP61]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP63]]) #[[ATTR4]]
-// CHECK5-NEXT:    br label [[OMP_IF_END26]]
-// CHECK5:       omp_if.end26:
-// CHECK5-NEXT:    [[TMP126:%.*]] = load i32, i32* [[A]], align 4
-// CHECK5-NEXT:    [[TMP127:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK5-NEXT:    call void @llvm.stackrestore(i8* [[TMP127]])
-// CHECK5-NEXT:    ret i32 [[TMP126]]
+// CHECK5-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 7
+// CHECK5-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to double**
+// CHECK5-NEXT:    store double* [[VLA1]], double** [[TMP122]], align 8
+// CHECK5-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 7
+// CHECK5-NEXT:    [[TMP124:%.*]] = bitcast i8** [[TMP123]] to double**
+// CHECK5-NEXT:    store double* [[VLA1]], double** [[TMP124]], align 8
+// CHECK5-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK5-NEXT:    store i64 [[TMP83]], i64* [[TMP125]], align 8
+// CHECK5-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 7
+// CHECK5-NEXT:    store i8* null, i8** [[TMP126]], align 8
+// CHECK5-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 8
+// CHECK5-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to %struct.TT**
+// CHECK5-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP128]], align 8
+// CHECK5-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 8
+// CHECK5-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to %struct.TT**
+// CHECK5-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP130]], align 8
+// CHECK5-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 8
+// CHECK5-NEXT:    store i8* null, i8** [[TMP131]], align 8
+// CHECK5-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 9
+// CHECK5-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP79]], i64* [[TMP133]], align 8
+// CHECK5-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 9
+// CHECK5-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP79]], i64* [[TMP135]], align 8
+// CHECK5-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 9
+// CHECK5-NEXT:    store i8* null, i8** [[TMP136]], align 8
+// CHECK5-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK5-NEXT:    [[KERNEL_ARGS24:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP140]], align 4
+// CHECK5-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 10, i32* [[TMP141]], align 4
+// CHECK5-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP137]], i8*** [[TMP142]], align 8
+// CHECK5-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP138]], i8*** [[TMP143]], align 8
+// CHECK5-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* [[TMP139]], i64** [[TMP144]], align 8
+// CHECK5-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP145]], align 8
+// CHECK5-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP146]], align 8
+// CHECK5-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP147]], align 8
+// CHECK5-NEXT:    [[TMP148:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]])
+// CHECK5-NEXT:    [[TMP149:%.*]] = icmp ne i32 [[TMP148]], 0
+// CHECK5-NEXT:    br i1 [[TMP149]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
+// CHECK5:       omp_offload.failed25:
+// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i64 [[TMP77]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP79]]) #[[ATTR4]]
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
+// CHECK5:       omp_offload.cont26:
+// CHECK5-NEXT:    br label [[OMP_IF_END28:%.*]]
+// CHECK5:       omp_if.else27:
+// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i64 [[TMP77]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP79]]) #[[ATTR4]]
+// CHECK5-NEXT:    br label [[OMP_IF_END28]]
+// CHECK5:       omp_if.end28:
+// CHECK5-NEXT:    [[TMP150:%.*]] = load i32, i32* [[A]], align 4
+// CHECK5-NEXT:    [[TMP151:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK5-NEXT:    call void @llvm.stackrestore(i8* [[TMP151]])
+// CHECK5-NEXT:    ret i32 [[TMP150]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96
@@ -3732,6 +4021,7 @@ int bar(int n){
 // CHECK5-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
 // CHECK5-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
 // CHECK5-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 8
+// CHECK5-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK5-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
 // CHECK5-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -3755,9 +4045,25 @@ int bar(int n){
 // CHECK5-NEXT:    store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !25
 // CHECK5-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !25
 // CHECK5-NEXT:    [[TMP10:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !25
-// CHECK5-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK5-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK5-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP11]], align 4, !noalias !25
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP12]], align 4, !noalias !25
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP13]], align 8, !noalias !25
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP14]], align 8, !noalias !25
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP15]], align 8, !noalias !25
+// CHECK5-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP16]], align 8, !noalias !25
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP17]], align 8, !noalias !25
+// CHECK5-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP18]], align 8, !noalias !25
+// CHECK5-NEXT:    [[TMP19:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK5-NEXT:    [[TMP20:%.*]] = icmp ne i32 [[TMP19]], 0
+// CHECK5-NEXT:    br i1 [[TMP20]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK5:       omp_offload.failed.i:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96() #[[ATTR4]]
 // CHECK5-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
@@ -4469,9 +4775,26 @@ int bar(int n){
 // CHECK5-NEXT:    [[TMP48:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
 // CHECK5-NEXT:    [[TOBOOL5:%.*]] = trunc i8 [[TMP48]] to i1
 // CHECK5-NEXT:    [[TMP49:%.*]] = select i1 [[TOBOOL5]], i32 0, i32 1
-// CHECK5-NEXT:    [[TMP50:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l214.region_id, i32 6, i8** [[TMP45]], i8** [[TMP46]], i64* [[TMP47]], i64* getelementptr inbounds ([6 x i64], [6 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP49]])
-// CHECK5-NEXT:    [[TMP51:%.*]] = icmp ne i32 [[TMP50]], 0
-// CHECK5-NEXT:    br i1 [[TMP51]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP50]], align 4
+// CHECK5-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 6, i32* [[TMP51]], align 4
+// CHECK5-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP45]], i8*** [[TMP52]], align 8
+// CHECK5-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP46]], i8*** [[TMP53]], align 8
+// CHECK5-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* [[TMP47]], i64** [[TMP54]], align 8
+// CHECK5-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([6 x i64], [6 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP55]], align 8
+// CHECK5-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP56]], align 8
+// CHECK5-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP57]], align 8
+// CHECK5-NEXT:    [[TMP58:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 [[TMP49]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l214.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP59:%.*]] = icmp ne i32 [[TMP58]], 0
+// CHECK5-NEXT:    br i1 [[TMP59]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l214(%struct.S1* [[THIS1]], i64 [[TMP7]], i64 2, i64 [[TMP2]], i16* [[VLA]], i64 [[TMP9]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -4481,15 +4804,15 @@ int bar(int n){
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l214(%struct.S1* [[THIS1]], i64 [[TMP7]], i64 2, i64 [[TMP2]], i16* [[VLA]], i64 [[TMP9]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_IF_END]]
 // CHECK5:       omp_if.end:
-// CHECK5-NEXT:    [[TMP52:%.*]] = mul nsw i64 1, [[TMP2]]
-// CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP52]]
+// CHECK5-NEXT:    [[TMP60:%.*]] = mul nsw i64 1, [[TMP2]]
+// CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP60]]
 // CHECK5-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
-// CHECK5-NEXT:    [[TMP53:%.*]] = load i16, i16* [[ARRAYIDX6]], align 2
-// CHECK5-NEXT:    [[CONV7:%.*]] = sext i16 [[TMP53]] to i32
-// CHECK5-NEXT:    [[TMP54:%.*]] = load i32, i32* [[B]], align 4
-// CHECK5-NEXT:    [[ADD8:%.*]] = add nsw i32 [[CONV7]], [[TMP54]]
-// CHECK5-NEXT:    [[TMP55:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK5-NEXT:    call void @llvm.stackrestore(i8* [[TMP55]])
+// CHECK5-NEXT:    [[TMP61:%.*]] = load i16, i16* [[ARRAYIDX6]], align 2
+// CHECK5-NEXT:    [[CONV7:%.*]] = sext i16 [[TMP61]] to i32
+// CHECK5-NEXT:    [[TMP62:%.*]] = load i32, i32* [[B]], align 4
+// CHECK5-NEXT:    [[ADD8:%.*]] = add nsw i32 [[CONV7]], [[TMP62]]
+// CHECK5-NEXT:    [[TMP63:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK5-NEXT:    call void @llvm.stackrestore(i8* [[TMP63]])
 // CHECK5-NEXT:    ret i32 [[ADD8]]
 //
 //
@@ -4561,9 +4884,26 @@ int bar(int n){
 // CHECK5-NEXT:    store i8* null, i8** [[TMP26]], align 8
 // CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK5-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK5-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK5-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK5-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK5-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK5-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK5-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK5-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -4573,8 +4913,8 @@ int bar(int n){
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_IF_END]]
 // CHECK5:       omp_if.end:
-// CHECK5-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP31]]
+// CHECK5-NEXT:    [[TMP39:%.*]] = load i32, i32* [[A]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP39]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -4630,9 +4970,26 @@ int bar(int n){
 // CHECK5-NEXT:    store i8* null, i8** [[TMP19]], align 8
 // CHECK5-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK5-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK5-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK5-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK5-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -4642,8 +4999,8 @@ int bar(int n){
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_IF_END]]
 // CHECK5:       omp_if.end:
-// CHECK5-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP24]]
+// CHECK5-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l214
@@ -5071,11 +5428,11 @@ int bar(int n){
 // CHECK7-NEXT:    [[DOTOFFLOAD_PTRS7:%.*]] = alloca [2 x i8*], align 4
 // CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS8:%.*]] = alloca [2 x i8*], align 4
 // CHECK7-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
-// CHECK7-NEXT:    [[A_CASTED11:%.*]] = alloca i32, align 4
+// CHECK7-NEXT:    [[A_CASTED12:%.*]] = alloca i32, align 4
 // CHECK7-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS14:%.*]] = alloca [10 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS15:%.*]] = alloca [10 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS16:%.*]] = alloca [10 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [10 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [10 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [10 x i8*], align 4
 // CHECK7-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [10 x i64], align 4
 // CHECK7-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK7-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
@@ -5137,176 +5494,227 @@ int bar(int n){
 // CHECK7-NEXT:    store i8* null, i8** [[TMP31]], align 4
 // CHECK7-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l108.region_id, i32 3, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK7-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK7-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK7-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 3, i32* [[TMP35]], align 4
+// CHECK7-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP32]], i8*** [[TMP36]], align 4
+// CHECK7-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP33]], i8*** [[TMP37]], align 4
+// CHECK7-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 4
+// CHECK7-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 4
+// CHECK7-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP40]], align 4
+// CHECK7-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP41]], align 4
+// CHECK7-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK7-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l108(i32 [[TMP12]], i32 [[TMP14]], i32 [[TMP16]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK7:       omp_offload.cont:
-// CHECK7-NEXT:    [[TMP36:%.*]] = load i32, i32* [[A]], align 4
-// CHECK7-NEXT:    store i32 [[TMP36]], i32* [[A_CASTED3]], align 4
-// CHECK7-NEXT:    [[TMP37:%.*]] = load i32, i32* [[A_CASTED3]], align 4
-// CHECK7-NEXT:    [[TMP38:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK7-NEXT:    [[TMP44:%.*]] = load i32, i32* [[A]], align 4
+// CHECK7-NEXT:    store i32 [[TMP44]], i32* [[A_CASTED3]], align 4
+// CHECK7-NEXT:    [[TMP45:%.*]] = load i32, i32* [[A_CASTED3]], align 4
+// CHECK7-NEXT:    [[TMP46:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK7-NEXT:    [[CONV5:%.*]] = bitcast i32* [[AA_CASTED4]] to i16*
-// CHECK7-NEXT:    store i16 [[TMP38]], i16* [[CONV5]], align 2
-// CHECK7-NEXT:    [[TMP39:%.*]] = load i32, i32* [[AA_CASTED4]], align 4
-// CHECK7-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK7-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP40]], 10
+// CHECK7-NEXT:    store i16 [[TMP46]], i16* [[CONV5]], align 2
+// CHECK7-NEXT:    [[TMP47:%.*]] = load i32, i32* [[AA_CASTED4]], align 4
+// CHECK7-NEXT:    [[TMP48:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK7-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP48]], 10
 // CHECK7-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK7:       omp_if.then:
-// CHECK7-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP37]], i32* [[TMP42]], align 4
-// CHECK7-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP37]], i32* [[TMP44]], align 4
-// CHECK7-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP45]], align 4
-// CHECK7-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 1
-// CHECK7-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP39]], i32* [[TMP47]], align 4
-// CHECK7-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 1
-// CHECK7-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP39]], i32* [[TMP49]], align 4
-// CHECK7-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 1
-// CHECK7-NEXT:    store i8* null, i8** [[TMP50]], align 4
-// CHECK7-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP53:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116.region_id, i32 2, i8** [[TMP51]], i8** [[TMP52]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK7-NEXT:    [[TMP54:%.*]] = icmp ne i32 [[TMP53]], 0
-// CHECK7-NEXT:    br i1 [[TMP54]], label [[OMP_OFFLOAD_FAILED9:%.*]], label [[OMP_OFFLOAD_CONT10:%.*]]
-// CHECK7:       omp_offload.failed9:
-// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i32 [[TMP37]], i32 [[TMP39]]) #[[ATTR4]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT10]]
-// CHECK7:       omp_offload.cont10:
+// CHECK7-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP45]], i32* [[TMP50]], align 4
+// CHECK7-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP45]], i32* [[TMP52]], align 4
+// CHECK7-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP53]], align 4
+// CHECK7-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 1
+// CHECK7-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP47]], i32* [[TMP55]], align 4
+// CHECK7-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 1
+// CHECK7-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP47]], i32* [[TMP57]], align 4
+// CHECK7-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS8]], i32 0, i32 1
+// CHECK7-NEXT:    store i8* null, i8** [[TMP58]], align 4
+// CHECK7-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
+// CHECK7-NEXT:    [[KERNEL_ARGS9:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP61]], align 4
+// CHECK7-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 2, i32* [[TMP62]], align 4
+// CHECK7-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP59]], i8*** [[TMP63]], align 4
+// CHECK7-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP60]], i8*** [[TMP64]], align 4
+// CHECK7-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP65]], align 4
+// CHECK7-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP66]], align 4
+// CHECK7-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP67]], align 4
+// CHECK7-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP68]], align 4
+// CHECK7-NEXT:    [[TMP69:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]])
+// CHECK7-NEXT:    [[TMP70:%.*]] = icmp ne i32 [[TMP69]], 0
+// CHECK7-NEXT:    br i1 [[TMP70]], label [[OMP_OFFLOAD_FAILED10:%.*]], label [[OMP_OFFLOAD_CONT11:%.*]]
+// CHECK7:       omp_offload.failed10:
+// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i32 [[TMP45]], i32 [[TMP47]]) #[[ATTR4]]
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT11]]
+// CHECK7:       omp_offload.cont11:
 // CHECK7-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK7:       omp_if.else:
-// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i32 [[TMP37]], i32 [[TMP39]]) #[[ATTR4]]
+// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l116(i32 [[TMP45]], i32 [[TMP47]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_IF_END]]
 // CHECK7:       omp_if.end:
-// CHECK7-NEXT:    [[TMP55:%.*]] = load i32, i32* [[A]], align 4
-// CHECK7-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK7-NEXT:    [[TMP56:%.*]] = load i32, i32* [[A]], align 4
-// CHECK7-NEXT:    store i32 [[TMP56]], i32* [[A_CASTED11]], align 4
-// CHECK7-NEXT:    [[TMP57:%.*]] = load i32, i32* [[A_CASTED11]], align 4
-// CHECK7-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK7-NEXT:    store i32 [[TMP58]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK7-NEXT:    [[TMP59:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK7-NEXT:    [[TMP60:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK7-NEXT:    [[CMP12:%.*]] = icmp sgt i32 [[TMP60]], 20
-// CHECK7-NEXT:    br i1 [[CMP12]], label [[OMP_IF_THEN13:%.*]], label [[OMP_IF_ELSE19:%.*]]
-// CHECK7:       omp_if.then13:
-// CHECK7-NEXT:    [[TMP61:%.*]] = mul nuw i32 [[TMP1]], 4
-// CHECK7-NEXT:    [[TMP62:%.*]] = sext i32 [[TMP61]] to i64
-// CHECK7-NEXT:    [[TMP63:%.*]] = mul nuw i32 5, [[TMP3]]
-// CHECK7-NEXT:    [[TMP64:%.*]] = mul nuw i32 [[TMP63]], 8
-// CHECK7-NEXT:    [[TMP65:%.*]] = sext i32 [[TMP64]] to i64
-// CHECK7-NEXT:    [[TMP66:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK7-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP66]], i8* align 4 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i32 80, i1 false)
-// CHECK7-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP57]], i32* [[TMP68]], align 4
-// CHECK7-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP57]], i32* [[TMP70]], align 4
-// CHECK7-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP71]], align 4
-// CHECK7-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 1
-// CHECK7-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to [10 x float]**
-// CHECK7-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP73]], align 4
-// CHECK7-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 1
-// CHECK7-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to [10 x float]**
-// CHECK7-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP75]], align 4
-// CHECK7-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 1
-// CHECK7-NEXT:    store i8* null, i8** [[TMP76]], align 4
-// CHECK7-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 2
-// CHECK7-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP1]], i32* [[TMP78]], align 4
-// CHECK7-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 2
-// CHECK7-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP1]], i32* [[TMP80]], align 4
-// CHECK7-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 2
-// CHECK7-NEXT:    store i8* null, i8** [[TMP81]], align 4
-// CHECK7-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 3
-// CHECK7-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to float**
-// CHECK7-NEXT:    store float* [[VLA]], float** [[TMP83]], align 4
-// CHECK7-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 3
-// CHECK7-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to float**
-// CHECK7-NEXT:    store float* [[VLA]], float** [[TMP85]], align 4
-// CHECK7-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK7-NEXT:    store i64 [[TMP62]], i64* [[TMP86]], align 4
-// CHECK7-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 3
+// CHECK7-NEXT:    [[TMP71:%.*]] = load i32, i32* [[A]], align 4
+// CHECK7-NEXT:    store i32 [[TMP71]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK7-NEXT:    [[TMP72:%.*]] = load i32, i32* [[A]], align 4
+// CHECK7-NEXT:    store i32 [[TMP72]], i32* [[A_CASTED12]], align 4
+// CHECK7-NEXT:    [[TMP73:%.*]] = load i32, i32* [[A_CASTED12]], align 4
+// CHECK7-NEXT:    [[TMP74:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK7-NEXT:    store i32 [[TMP74]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK7-NEXT:    [[TMP75:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK7-NEXT:    [[TMP76:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK7-NEXT:    [[CMP13:%.*]] = icmp sgt i32 [[TMP76]], 20
+// CHECK7-NEXT:    br i1 [[CMP13]], label [[OMP_IF_THEN14:%.*]], label [[OMP_IF_ELSE21:%.*]]
+// CHECK7:       omp_if.then14:
+// CHECK7-NEXT:    [[TMP77:%.*]] = mul nuw i32 [[TMP1]], 4
+// CHECK7-NEXT:    [[TMP78:%.*]] = sext i32 [[TMP77]] to i64
+// CHECK7-NEXT:    [[TMP79:%.*]] = mul nuw i32 5, [[TMP3]]
+// CHECK7-NEXT:    [[TMP80:%.*]] = mul nuw i32 [[TMP79]], 8
+// CHECK7-NEXT:    [[TMP81:%.*]] = sext i32 [[TMP80]] to i64
+// CHECK7-NEXT:    [[TMP82:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK7-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP82]], i8* align 4 bitcast ([10 x i64]* @.offload_sizes.8 to i8*), i32 80, i1 false)
+// CHECK7-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP73]], i32* [[TMP84]], align 4
+// CHECK7-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP73]], i32* [[TMP86]], align 4
+// CHECK7-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 0
 // CHECK7-NEXT:    store i8* null, i8** [[TMP87]], align 4
-// CHECK7-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 4
-// CHECK7-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to [5 x [10 x double]]**
-// CHECK7-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP89]], align 4
-// CHECK7-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 4
-// CHECK7-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to [5 x [10 x double]]**
-// CHECK7-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP91]], align 4
-// CHECK7-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 4
+// CHECK7-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 1
+// CHECK7-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to [10 x float]**
+// CHECK7-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP89]], align 4
+// CHECK7-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 1
+// CHECK7-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to [10 x float]**
+// CHECK7-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP91]], align 4
+// CHECK7-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 1
 // CHECK7-NEXT:    store i8* null, i8** [[TMP92]], align 4
-// CHECK7-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 5
+// CHECK7-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 2
 // CHECK7-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32*
-// CHECK7-NEXT:    store i32 5, i32* [[TMP94]], align 4
-// CHECK7-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 5
+// CHECK7-NEXT:    store i32 [[TMP1]], i32* [[TMP94]], align 4
+// CHECK7-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 2
 // CHECK7-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32*
-// CHECK7-NEXT:    store i32 5, i32* [[TMP96]], align 4
-// CHECK7-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 5
+// CHECK7-NEXT:    store i32 [[TMP1]], i32* [[TMP96]], align 4
+// CHECK7-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 2
 // CHECK7-NEXT:    store i8* null, i8** [[TMP97]], align 4
-// CHECK7-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 6
-// CHECK7-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP3]], i32* [[TMP99]], align 4
-// CHECK7-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 6
-// CHECK7-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP3]], i32* [[TMP101]], align 4
-// CHECK7-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 6
-// CHECK7-NEXT:    store i8* null, i8** [[TMP102]], align 4
-// CHECK7-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 7
-// CHECK7-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to double**
-// CHECK7-NEXT:    store double* [[VLA1]], double** [[TMP104]], align 4
-// CHECK7-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 7
-// CHECK7-NEXT:    [[TMP106:%.*]] = bitcast i8** [[TMP105]] to double**
-// CHECK7-NEXT:    store double* [[VLA1]], double** [[TMP106]], align 4
-// CHECK7-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK7-NEXT:    store i64 [[TMP65]], i64* [[TMP107]], align 4
-// CHECK7-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 7
+// CHECK7-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 3
+// CHECK7-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to float**
+// CHECK7-NEXT:    store float* [[VLA]], float** [[TMP99]], align 4
+// CHECK7-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 3
+// CHECK7-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to float**
+// CHECK7-NEXT:    store float* [[VLA]], float** [[TMP101]], align 4
+// CHECK7-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK7-NEXT:    store i64 [[TMP78]], i64* [[TMP102]], align 4
+// CHECK7-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 3
+// CHECK7-NEXT:    store i8* null, i8** [[TMP103]], align 4
+// CHECK7-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 4
+// CHECK7-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to [5 x [10 x double]]**
+// CHECK7-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP105]], align 4
+// CHECK7-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 4
+// CHECK7-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to [5 x [10 x double]]**
+// CHECK7-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP107]], align 4
+// CHECK7-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 4
 // CHECK7-NEXT:    store i8* null, i8** [[TMP108]], align 4
-// CHECK7-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 8
-// CHECK7-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to %struct.TT**
-// CHECK7-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP110]], align 4
-// CHECK7-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 8
-// CHECK7-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to %struct.TT**
-// CHECK7-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP112]], align 4
-// CHECK7-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 8
+// CHECK7-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 5
+// CHECK7-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32*
+// CHECK7-NEXT:    store i32 5, i32* [[TMP110]], align 4
+// CHECK7-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 5
+// CHECK7-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32*
+// CHECK7-NEXT:    store i32 5, i32* [[TMP112]], align 4
+// CHECK7-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 5
 // CHECK7-NEXT:    store i8* null, i8** [[TMP113]], align 4
-// CHECK7-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 9
+// CHECK7-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 6
 // CHECK7-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP59]], i32* [[TMP115]], align 4
-// CHECK7-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 9
+// CHECK7-NEXT:    store i32 [[TMP3]], i32* [[TMP115]], align 4
+// CHECK7-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 6
 // CHECK7-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP59]], i32* [[TMP117]], align 4
-// CHECK7-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i32 0, i32 9
+// CHECK7-NEXT:    store i32 [[TMP3]], i32* [[TMP117]], align 4
+// CHECK7-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 6
 // CHECK7-NEXT:    store i8* null, i8** [[TMP118]], align 4
-// CHECK7-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP122:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140.region_id, i32 10, i8** [[TMP119]], i8** [[TMP120]], i64* [[TMP121]], i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK7-NEXT:    [[TMP123:%.*]] = icmp ne i32 [[TMP122]], 0
-// CHECK7-NEXT:    br i1 [[TMP123]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK7:       omp_offload.failed17:
-// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i32 [[TMP57]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP59]]) #[[ATTR4]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK7:       omp_offload.cont18:
-// CHECK7-NEXT:    br label [[OMP_IF_END20:%.*]]
-// CHECK7:       omp_if.else19:
-// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i32 [[TMP57]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP59]]) #[[ATTR4]]
-// CHECK7-NEXT:    br label [[OMP_IF_END20]]
-// CHECK7:       omp_if.end20:
-// CHECK7-NEXT:    [[TMP124:%.*]] = load i32, i32* [[A]], align 4
-// CHECK7-NEXT:    [[TMP125:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK7-NEXT:    call void @llvm.stackrestore(i8* [[TMP125]])
-// CHECK7-NEXT:    ret i32 [[TMP124]]
+// CHECK7-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 7
+// CHECK7-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to double**
+// CHECK7-NEXT:    store double* [[VLA1]], double** [[TMP120]], align 4
+// CHECK7-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 7
+// CHECK7-NEXT:    [[TMP122:%.*]] = bitcast i8** [[TMP121]] to double**
+// CHECK7-NEXT:    store double* [[VLA1]], double** [[TMP122]], align 4
+// CHECK7-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK7-NEXT:    store i64 [[TMP81]], i64* [[TMP123]], align 4
+// CHECK7-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 7
+// CHECK7-NEXT:    store i8* null, i8** [[TMP124]], align 4
+// CHECK7-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 8
+// CHECK7-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to %struct.TT**
+// CHECK7-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP126]], align 4
+// CHECK7-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 8
+// CHECK7-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to %struct.TT**
+// CHECK7-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP128]], align 4
+// CHECK7-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 8
+// CHECK7-NEXT:    store i8* null, i8** [[TMP129]], align 4
+// CHECK7-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 9
+// CHECK7-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP75]], i32* [[TMP131]], align 4
+// CHECK7-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 9
+// CHECK7-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP75]], i32* [[TMP133]], align 4
+// CHECK7-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 9
+// CHECK7-NEXT:    store i8* null, i8** [[TMP134]], align 4
+// CHECK7-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK7-NEXT:    [[KERNEL_ARGS18:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP138]], align 4
+// CHECK7-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 10, i32* [[TMP139]], align 4
+// CHECK7-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP135]], i8*** [[TMP140]], align 4
+// CHECK7-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP136]], i8*** [[TMP141]], align 4
+// CHECK7-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* [[TMP137]], i64** [[TMP142]], align 4
+// CHECK7-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP143]], align 4
+// CHECK7-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP144]], align 4
+// CHECK7-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP145]], align 4
+// CHECK7-NEXT:    [[TMP146:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]])
+// CHECK7-NEXT:    [[TMP147:%.*]] = icmp ne i32 [[TMP146]], 0
+// CHECK7-NEXT:    br i1 [[TMP147]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
+// CHECK7:       omp_offload.failed19:
+// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i32 [[TMP73]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP75]]) #[[ATTR4]]
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
+// CHECK7:       omp_offload.cont20:
+// CHECK7-NEXT:    br label [[OMP_IF_END22:%.*]]
+// CHECK7:       omp_if.else21:
+// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l140(i32 [[TMP73]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP75]]) #[[ATTR4]]
+// CHECK7-NEXT:    br label [[OMP_IF_END22]]
+// CHECK7:       omp_if.end22:
+// CHECK7-NEXT:    [[TMP148:%.*]] = load i32, i32* [[A]], align 4
+// CHECK7-NEXT:    [[TMP149:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK7-NEXT:    call void @llvm.stackrestore(i8* [[TMP149]])
+// CHECK7-NEXT:    ret i32 [[TMP148]]
 //
 //
 // CHECK7-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96
@@ -5392,6 +5800,7 @@ int bar(int n){
 // CHECK7-NEXT:    [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 4
 // CHECK7-NEXT:    [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 4
 // CHECK7-NEXT:    [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 4
+// CHECK7-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK7-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK7-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 4
 // CHECK7-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -5415,9 +5824,25 @@ int bar(int n){
 // CHECK7-NEXT:    store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 4, !noalias !26
 // CHECK7-NEXT:    store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !26
 // CHECK7-NEXT:    [[TMP10:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 4, !noalias !26
-// CHECK7-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK7-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK7-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK7-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP11]], align 4, !noalias !26
+// CHECK7-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 0, i32* [[TMP12]], align 4, !noalias !26
+// CHECK7-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP13]], align 4, !noalias !26
+// CHECK7-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP14]], align 4, !noalias !26
+// CHECK7-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* null, i64** [[TMP15]], align 4, !noalias !26
+// CHECK7-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* null, i64** [[TMP16]], align 4, !noalias !26
+// CHECK7-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP17]], align 4, !noalias !26
+// CHECK7-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP18]], align 4, !noalias !26
+// CHECK7-NEXT:    [[TMP19:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK7-NEXT:    [[TMP20:%.*]] = icmp ne i32 [[TMP19]], 0
+// CHECK7-NEXT:    br i1 [[TMP20]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK7:       omp_offload.failed.i:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l96() #[[ATTR4]]
 // CHECK7-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
@@ -6108,9 +6533,26 @@ int bar(int n){
 // CHECK7-NEXT:    [[TMP48:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
 // CHECK7-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP48]] to i1
 // CHECK7-NEXT:    [[TMP49:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
-// CHECK7-NEXT:    [[TMP50:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l214.region_id, i32 6, i8** [[TMP45]], i8** [[TMP46]], i64* [[TMP47]], i64* getelementptr inbounds ([6 x i64], [6 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP49]])
-// CHECK7-NEXT:    [[TMP51:%.*]] = icmp ne i32 [[TMP50]], 0
-// CHECK7-NEXT:    br i1 [[TMP51]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP50]], align 4
+// CHECK7-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 6, i32* [[TMP51]], align 4
+// CHECK7-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP45]], i8*** [[TMP52]], align 4
+// CHECK7-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP46]], i8*** [[TMP53]], align 4
+// CHECK7-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* [[TMP47]], i64** [[TMP54]], align 4
+// CHECK7-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([6 x i64], [6 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP55]], align 4
+// CHECK7-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP56]], align 4
+// CHECK7-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP57]], align 4
+// CHECK7-NEXT:    [[TMP58:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 [[TMP49]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l214.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP59:%.*]] = icmp ne i32 [[TMP58]], 0
+// CHECK7-NEXT:    br i1 [[TMP59]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l214(%struct.S1* [[THIS1]], i32 [[TMP6]], i32 2, i32 [[TMP1]], i16* [[VLA]], i32 [[TMP8]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -6120,15 +6562,15 @@ int bar(int n){
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l214(%struct.S1* [[THIS1]], i32 [[TMP6]], i32 2, i32 [[TMP1]], i16* [[VLA]], i32 [[TMP8]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_IF_END]]
 // CHECK7:       omp_if.end:
-// CHECK7-NEXT:    [[TMP52:%.*]] = mul nsw i32 1, [[TMP1]]
-// CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP52]]
+// CHECK7-NEXT:    [[TMP60:%.*]] = mul nsw i32 1, [[TMP1]]
+// CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP60]]
 // CHECK7-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
-// CHECK7-NEXT:    [[TMP53:%.*]] = load i16, i16* [[ARRAYIDX5]], align 2
-// CHECK7-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP53]] to i32
-// CHECK7-NEXT:    [[TMP54:%.*]] = load i32, i32* [[B]], align 4
-// CHECK7-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], [[TMP54]]
-// CHECK7-NEXT:    [[TMP55:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK7-NEXT:    call void @llvm.stackrestore(i8* [[TMP55]])
+// CHECK7-NEXT:    [[TMP61:%.*]] = load i16, i16* [[ARRAYIDX5]], align 2
+// CHECK7-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP61]] to i32
+// CHECK7-NEXT:    [[TMP62:%.*]] = load i32, i32* [[B]], align 4
+// CHECK7-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], [[TMP62]]
+// CHECK7-NEXT:    [[TMP63:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK7-NEXT:    call void @llvm.stackrestore(i8* [[TMP63]])
 // CHECK7-NEXT:    ret i32 [[ADD7]]
 //
 //
@@ -6199,9 +6641,26 @@ int bar(int n){
 // CHECK7-NEXT:    store i8* null, i8** [[TMP26]], align 4
 // CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK7-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK7-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK7-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK7-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK7-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK7-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK7-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK7-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK7-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK7-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK7-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -6211,8 +6670,8 @@ int bar(int n){
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l195(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_IF_END]]
 // CHECK7:       omp_if.end:
-// CHECK7-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
-// CHECK7-NEXT:    ret i32 [[TMP31]]
+// CHECK7-NEXT:    [[TMP39:%.*]] = load i32, i32* [[A]], align 4
+// CHECK7-NEXT:    ret i32 [[TMP39]]
 //
 //
 // CHECK7-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -6267,9 +6726,26 @@ int bar(int n){
 // CHECK7-NEXT:    store i8* null, i8** [[TMP19]], align 4
 // CHECK7-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK7-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK7-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK7-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK7-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 4
+// CHECK7-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 4
+// CHECK7-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK7-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK7-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK7-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -6279,8 +6755,8 @@ int bar(int n){
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l178(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_IF_END]]
 // CHECK7:       omp_if.end:
-// CHECK7-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK7-NEXT:    ret i32 [[TMP24]]
+// CHECK7-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK7-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK7-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l214

diff --git a/clang/test/OpenMP/target_parallel_for_simd_depend_codegen.cpp b/clang/test/OpenMP/target_parallel_for_simd_depend_codegen.cpp
index 91f3bb7..60c77d9 100644 (file)
--- a/clang/test/OpenMP/target_parallel_for_simd_depend_codegen.cpp
+++ b/clang/test/OpenMP/target_parallel_for_simd_depend_codegen.cpp
@@ -182,7 +182,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 0
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // CHECK:       [[FAIL]]
@@ -200,7 +200,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 2
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 2, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SZT]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 0, i32 0, i8* null, i32 0, i8* null)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 1, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 
 // CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]

diff --git a/clang/test/OpenMP/target_parallel_for_simd_uses_allocators_codegen.cpp b/clang/test/OpenMP/target_parallel_for_simd_uses_allocators_codegen.cpp
index 6c393fd..21e32fd 100644 (file)
--- a/clang/test/OpenMP/target_parallel_for_simd_uses_allocators_codegen.cpp
+++ b/clang/test/OpenMP/target_parallel_for_simd_uses_allocators_codegen.cpp
@@ -65,7 +65,7 @@ void foo() {
   omp_alloctrait_t traits[10];
   omp_allocator_handle_t my_allocator;
 
-// CHECK: [[RES:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @.[[TGT_REGION:.+]].region_id, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZES]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPTYPES]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
+// CHECK: [[RES:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i32 0, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
 // CHECK: [[CMP:%.+]] = icmp ne i32 [[RES]], 0
 // CHECK: br i1 [[CMP]], label %[[FAILED:.+]], label %[[DONE:.+]]
 // CHECK: [[FAILED]]:

diff --git a/clang/test/OpenMP/target_parallel_for_uses_allocators_codegen.cpp b/clang/test/OpenMP/target_parallel_for_uses_allocators_codegen.cpp
index 85c3b31..76370f4 100644 (file)
--- a/clang/test/OpenMP/target_parallel_for_uses_allocators_codegen.cpp
+++ b/clang/test/OpenMP/target_parallel_for_uses_allocators_codegen.cpp
@@ -65,7 +65,7 @@ void foo() {
   omp_alloctrait_t traits[10];
   omp_allocator_handle_t my_allocator;
 
-// CHECK: [[RES:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @.[[TGT_REGION:.+]].region_id, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZES]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPTYPES]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
+// CHECK: [[RES:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i32 0, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
 // CHECK: [[CMP:%.+]] = icmp ne i32 [[RES]], 0
 // CHECK: br i1 [[CMP]], label %[[FAILED:.+]], label %[[DONE:.+]]
 // CHECK: [[FAILED]]:

diff --git a/clang/test/OpenMP/target_parallel_if_codegen.cpp b/clang/test/OpenMP/target_parallel_if_codegen.cpp
index 548154c..71b1833 100644 (file)
--- a/clang/test/OpenMP/target_parallel_if_codegen.cpp
+++ b/clang/test/OpenMP/target_parallel_if_codegen.cpp
@@ -274,65 +274,99 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP22:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
 // CHECK1-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP22]] to i1
 // CHECK1-NEXT:    [[TMP23:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP24:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP23]])
-// CHECK1-NEXT:    [[TMP25:%.*]] = icmp ne i32 [[TMP24]], 0
-// CHECK1-NEXT:    br i1 [[TMP25]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 1, i32 [[TMP23]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121(%struct.S1* [[THIS1]], i64 [[TMP2]], i64 [[TMP4]]) #[[ATTR3:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP26:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP6:%.*]] = icmp sgt i32 [[TMP26]], 5
+// CHECK1-NEXT:    [[TMP34:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP6:%.*]] = icmp sgt i32 [[TMP34]], 5
 // CHECK1-NEXT:    [[FROMBOOL7:%.*]] = zext i1 [[CMP6]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL7]], i8* [[DOTCAPTURE_EXPR_5]], align 1
-// CHECK1-NEXT:    [[TMP27:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_5]], align 1
-// CHECK1-NEXT:    [[TOBOOL8:%.*]] = trunc i8 [[TMP27]] to i1
+// CHECK1-NEXT:    [[TMP35:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_5]], align 1
+// CHECK1-NEXT:    [[TOBOOL8:%.*]] = trunc i8 [[TMP35]] to i1
 // CHECK1-NEXT:    [[CONV10:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED9]] to i8*
 // CHECK1-NEXT:    [[FROMBOOL11:%.*]] = zext i1 [[TOBOOL8]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL11]], i8* [[CONV10]], align 1
-// CHECK1-NEXT:    [[TMP28:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED9]], align 8
-// CHECK1-NEXT:    [[TMP29:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP12:%.*]] = icmp sgt i32 [[TMP29]], 4
+// CHECK1-NEXT:    [[TMP36:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED9]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP12:%.*]] = icmp sgt i32 [[TMP37]], 4
 // CHECK1-NEXT:    br i1 [[CMP12]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK1:       omp_if.then:
 // CHECK1-NEXT:    [[A13:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to %struct.S1**
-// CHECK1-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP31]], align 8
-// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP33:%.*]] = bitcast i8** [[TMP32]] to double**
-// CHECK1-NEXT:    store double* [[A13]], double** [[TMP33]], align 8
-// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP34]], align 8
-// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP28]], i64* [[TMP36]], align 8
-// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP28]], i64* [[TMP38]], align 8
-// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP39]], align 8
-// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP42:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_5]], align 1
-// CHECK1-NEXT:    [[TOBOOL17:%.*]] = trunc i8 [[TMP42]] to i1
-// CHECK1-NEXT:    [[TMP43:%.*]] = select i1 [[TOBOOL17]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, i32 2, i8** [[TMP40]], i8** [[TMP41]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP43]])
-// CHECK1-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
-// CHECK1-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED18:%.*]], label [[OMP_OFFLOAD_CONT19:%.*]]
-// CHECK1:       omp_offload.failed18:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]], i64 [[TMP28]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT19]]
-// CHECK1:       omp_offload.cont19:
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to %struct.S1**
+// CHECK1-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to double**
+// CHECK1-NEXT:    store double* [[A13]], double** [[TMP41]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP36]], i64* [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP36]], i64* [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS16]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS14]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS15]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP50:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_5]], align 1
+// CHECK1-NEXT:    [[TOBOOL17:%.*]] = trunc i8 [[TMP50]] to i1
+// CHECK1-NEXT:    [[TMP51:%.*]] = select i1 [[TOBOOL17]], i32 0, i32 1
+// CHECK1-NEXT:    [[KERNEL_ARGS18:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP52]], align 4
+// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP53]], align 4
+// CHECK1-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP48]], i8*** [[TMP54]], align 8
+// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP49]], i8*** [[TMP55]], align 8
+// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP56]], align 8
+// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP57]], align 8
+// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP58]], align 8
+// CHECK1-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP59]], align 8
+// CHECK1-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 [[TMP51]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]])
+// CHECK1-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
+// CHECK1-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
+// CHECK1:       omp_offload.failed19:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]], i64 [[TMP36]]) #[[ATTR3]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
+// CHECK1:       omp_offload.cont20:
 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK1:       omp_if.else:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]], i64 [[TMP28]]) #[[ATTR3]]
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]], i64 [[TMP36]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[A20:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP46:%.*]] = load double, double* [[A20]], align 8
-// CHECK1-NEXT:    [[CONV21:%.*]] = fptosi double [[TMP46]] to i32
-// CHECK1-NEXT:    ret i32 [[CONV21]]
+// CHECK1-NEXT:    [[A21:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP62:%.*]] = load double, double* [[A21]], align 8
+// CHECK1-NEXT:    [[CONV22:%.*]] = fptosi double [[TMP62]] to i32
+// CHECK1-NEXT:    ret i32 [[CONV22]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZL7fstatici
@@ -372,9 +406,26 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP11:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
 // CHECK1-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP11]] to i1
 // CHECK1-NEXT:    [[TMP12:%.*]] = select i1 [[TOBOOL3]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, i32 1, i8** [[TMP9]], i8** [[TMP10]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP12]])
-// CHECK1-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK1-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP13]], align 4
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP14]], align 4
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP9]], i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP10]], i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 [[TMP12]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP22:%.*]] = icmp ne i32 [[TMP21]], 0
+// CHECK1-NEXT:    br i1 [[TMP22]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104(i64 [[TMP2]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -384,25 +435,42 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104(i64 [[TMP2]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP15:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[SUB:%.*]] = sub nsw i32 [[TMP15]], 2
+// CHECK1-NEXT:    [[TMP23:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[SUB:%.*]] = sub nsw i32 [[TMP23]], 2
 // CHECK1-NEXT:    [[CMP4:%.*]] = icmp sgt i32 [[SUB]], 2
-// CHECK1-NEXT:    br i1 [[CMP4]], label [[OMP_IF_THEN5:%.*]], label [[OMP_IF_ELSE8:%.*]]
+// CHECK1-NEXT:    br i1 [[CMP4]], label [[OMP_IF_THEN5:%.*]], label [[OMP_IF_ELSE9:%.*]]
 // CHECK1:       omp_if.then5:
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK1:       omp_offload.failed6:
+// CHECK1-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK1:       omp_offload.failed7:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108() #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK1:       omp_offload.cont7:
-// CHECK1-NEXT:    br label [[OMP_IF_END9:%.*]]
-// CHECK1:       omp_if.else8:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK1:       omp_offload.cont8:
+// CHECK1-NEXT:    br label [[OMP_IF_END10:%.*]]
+// CHECK1:       omp_if.else9:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108() #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_IF_END9]]
-// CHECK1:       omp_if.end9:
-// CHECK1-NEXT:    [[TMP18:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP18]], 1
+// CHECK1-NEXT:    br label [[OMP_IF_END10]]
+// CHECK1:       omp_if.end10:
+// CHECK1-NEXT:    [[TMP34:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP34]], 1
 // CHECK1-NEXT:    ret i32 [[ADD]]
 //
 //
@@ -437,49 +505,83 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP6]], align 8
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l87.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.9, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.10, i32 0, i32 0), i8** null, i8** null, i32 1, i32 1)
-// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.9, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.10, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l87.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l87(i64 [[TMP1]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i16 1, i16* [[B]], align 2
-// CHECK1-NEXT:    [[TMP11:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP19:%.*]] = load i32, i32* [[A]], align 4
 // CHECK1-NEXT:    [[CONV2:%.*]] = bitcast i64* [[A_CASTED1]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP11]], i32* [[CONV2]], align 4
-// CHECK1-NEXT:    [[TMP12:%.*]] = load i64, i64* [[A_CASTED1]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = load i16, i16* [[B]], align 2
+// CHECK1-NEXT:    store i32 [[TMP19]], i32* [[CONV2]], align 4
+// CHECK1-NEXT:    [[TMP20:%.*]] = load i64, i64* [[A_CASTED1]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i16, i16* [[B]], align 2
 // CHECK1-NEXT:    [[CONV3:%.*]] = bitcast i64* [[B_CASTED]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP13]], i16* [[CONV3]], align 2
-// CHECK1-NEXT:    [[TMP14:%.*]] = load i64, i64* [[B_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP16]], align 8
-// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP18]], align 8
-// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP21]], align 8
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP23]], align 8
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP24]], align 8
-// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, i32 2, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK1-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK1-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i64 [[TMP12]], i64 [[TMP14]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[TMP29:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP29]]
+// CHECK1-NEXT:    store i16 [[TMP21]], i16* [[CONV3]], align 2
+// CHECK1-NEXT:    [[TMP22:%.*]] = load i64, i64* [[B_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP20]], i64* [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP20]], i64* [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP22]], i64* [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP22]], i64* [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP35]], align 4
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP36]], align 4
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP33]], i8*** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP41]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
+// CHECK1-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i64 [[TMP20]], i64 [[TMP22]]) #[[ATTR3]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[TMP45:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP45]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121
@@ -817,65 +919,99 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP22:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
 // CHECK3-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP22]] to i1
 // CHECK3-NEXT:    [[TMP23:%.*]] = select i1 [[TOBOOL3]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP24:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP23]])
-// CHECK3-NEXT:    [[TMP25:%.*]] = icmp ne i32 [[TMP24]], 0
-// CHECK3-NEXT:    br i1 [[TMP25]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP20]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 1, i32 [[TMP23]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121(%struct.S1* [[THIS1]], i32 [[TMP2]], i32 [[TMP4]]) #[[ATTR3:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP26:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP5:%.*]] = icmp sgt i32 [[TMP26]], 5
+// CHECK3-NEXT:    [[TMP34:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP5:%.*]] = icmp sgt i32 [[TMP34]], 5
 // CHECK3-NEXT:    [[FROMBOOL6:%.*]] = zext i1 [[CMP5]] to i8
 // CHECK3-NEXT:    store i8 [[FROMBOOL6]], i8* [[DOTCAPTURE_EXPR_4]], align 1
-// CHECK3-NEXT:    [[TMP27:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_4]], align 1
-// CHECK3-NEXT:    [[TOBOOL7:%.*]] = trunc i8 [[TMP27]] to i1
+// CHECK3-NEXT:    [[TMP35:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_4]], align 1
+// CHECK3-NEXT:    [[TOBOOL7:%.*]] = trunc i8 [[TMP35]] to i1
 // CHECK3-NEXT:    [[CONV9:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR__CASTED8]] to i8*
 // CHECK3-NEXT:    [[FROMBOOL10:%.*]] = zext i1 [[TOBOOL7]] to i8
 // CHECK3-NEXT:    store i8 [[FROMBOOL10]], i8* [[CONV9]], align 1
-// CHECK3-NEXT:    [[TMP28:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED8]], align 4
-// CHECK3-NEXT:    [[TMP29:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP11:%.*]] = icmp sgt i32 [[TMP29]], 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED8]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP11:%.*]] = icmp sgt i32 [[TMP37]], 4
 // CHECK3-NEXT:    br i1 [[CMP11]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK3:       omp_if.then:
 // CHECK3-NEXT:    [[A12:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to %struct.S1**
-// CHECK3-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP31]], align 4
-// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP33:%.*]] = bitcast i8** [[TMP32]] to double**
-// CHECK3-NEXT:    store double* [[A12]], double** [[TMP33]], align 4
-// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP34]], align 4
-// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP28]], i32* [[TMP36]], align 4
-// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP28]], i32* [[TMP38]], align 4
-// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP39]], align 4
-// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP42:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_4]], align 1
-// CHECK3-NEXT:    [[TOBOOL16:%.*]] = trunc i8 [[TMP42]] to i1
-// CHECK3-NEXT:    [[TMP43:%.*]] = select i1 [[TOBOOL16]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, i32 2, i8** [[TMP40]], i8** [[TMP41]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP43]])
-// CHECK3-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
-// CHECK3-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK3:       omp_offload.failed17:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]], i32 [[TMP28]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK3:       omp_offload.cont18:
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to %struct.S1**
+// CHECK3-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP39]], align 4
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to double**
+// CHECK3-NEXT:    store double* [[A12]], double** [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP36]], i32* [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP36]], i32* [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP50:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_4]], align 1
+// CHECK3-NEXT:    [[TOBOOL16:%.*]] = trunc i8 [[TMP50]] to i1
+// CHECK3-NEXT:    [[TMP51:%.*]] = select i1 [[TOBOOL16]], i32 0, i32 1
+// CHECK3-NEXT:    [[KERNEL_ARGS17:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP52]], align 4
+// CHECK3-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP53]], align 4
+// CHECK3-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP48]], i8*** [[TMP54]], align 4
+// CHECK3-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP49]], i8*** [[TMP55]], align 4
+// CHECK3-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP56]], align 4
+// CHECK3-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP57]], align 4
+// CHECK3-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP58]], align 4
+// CHECK3-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP59]], align 4
+// CHECK3-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 [[TMP51]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS17]])
+// CHECK3-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
+// CHECK3-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED18:%.*]], label [[OMP_OFFLOAD_CONT19:%.*]]
+// CHECK3:       omp_offload.failed18:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]], i32 [[TMP36]]) #[[ATTR3]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT19]]
+// CHECK3:       omp_offload.cont19:
 // CHECK3-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK3:       omp_if.else:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]], i32 [[TMP28]]) #[[ATTR3]]
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]], i32 [[TMP36]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[A19:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP46:%.*]] = load double, double* [[A19]], align 4
-// CHECK3-NEXT:    [[CONV20:%.*]] = fptosi double [[TMP46]] to i32
-// CHECK3-NEXT:    ret i32 [[CONV20]]
+// CHECK3-NEXT:    [[A20:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP62:%.*]] = load double, double* [[A20]], align 4
+// CHECK3-NEXT:    [[CONV21:%.*]] = fptosi double [[TMP62]] to i32
+// CHECK3-NEXT:    ret i32 [[CONV21]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZL7fstatici
@@ -915,9 +1051,26 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP11:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
 // CHECK3-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP11]] to i1
 // CHECK3-NEXT:    [[TMP12:%.*]] = select i1 [[TOBOOL3]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP13:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, i32 1, i8** [[TMP9]], i8** [[TMP10]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP12]])
-// CHECK3-NEXT:    [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
-// CHECK3-NEXT:    br i1 [[TMP14]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP9]], i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP10]], i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP19]], align 4
+// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 [[TMP12]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP22:%.*]] = icmp ne i32 [[TMP21]], 0
+// CHECK3-NEXT:    br i1 [[TMP22]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104(i32 [[TMP2]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -927,25 +1080,42 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104(i32 [[TMP2]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP15:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[SUB:%.*]] = sub nsw i32 [[TMP15]], 2
+// CHECK3-NEXT:    [[TMP23:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[SUB:%.*]] = sub nsw i32 [[TMP23]], 2
 // CHECK3-NEXT:    [[CMP4:%.*]] = icmp sgt i32 [[SUB]], 2
-// CHECK3-NEXT:    br i1 [[CMP4]], label [[OMP_IF_THEN5:%.*]], label [[OMP_IF_ELSE8:%.*]]
+// CHECK3-NEXT:    br i1 [[CMP4]], label [[OMP_IF_THEN5:%.*]], label [[OMP_IF_ELSE9:%.*]]
 // CHECK3:       omp_if.then5:
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK3:       omp_offload.failed6:
+// CHECK3-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK3:       omp_offload.failed7:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108() #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK3:       omp_offload.cont7:
-// CHECK3-NEXT:    br label [[OMP_IF_END9:%.*]]
-// CHECK3:       omp_if.else8:
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK3:       omp_offload.cont8:
+// CHECK3-NEXT:    br label [[OMP_IF_END10:%.*]]
+// CHECK3:       omp_if.else9:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108() #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_IF_END9]]
-// CHECK3:       omp_if.end9:
-// CHECK3-NEXT:    [[TMP18:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP18]], 1
+// CHECK3-NEXT:    br label [[OMP_IF_END10]]
+// CHECK3:       omp_if.end10:
+// CHECK3-NEXT:    [[TMP34:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP34]], 1
 // CHECK3-NEXT:    ret i32 [[ADD]]
 //
 //
@@ -979,48 +1149,82 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP6]], align 4
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l87.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.9, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.10, i32 0, i32 0), i8** null, i8** null, i32 1, i32 1)
-// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.9, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.10, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l87.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK3-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l87(i32 [[TMP1]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i16 1, i16* [[B]], align 2
-// CHECK3-NEXT:    [[TMP11:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP11]], i32* [[A_CASTED1]], align 4
-// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[A_CASTED1]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = load i16, i16* [[B]], align 2
+// CHECK3-NEXT:    [[TMP19:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP19]], i32* [[A_CASTED1]], align 4
+// CHECK3-NEXT:    [[TMP20:%.*]] = load i32, i32* [[A_CASTED1]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = load i16, i16* [[B]], align 2
 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[B_CASTED]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP13]], i16* [[CONV]], align 2
-// CHECK3-NEXT:    [[TMP14:%.*]] = load i32, i32* [[B_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP16]], align 4
-// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP18]], align 4
-// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP14]], i32* [[TMP21]], align 4
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP14]], i32* [[TMP23]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP24]], align 4
-// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, i32 2, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
-// CHECK3-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK3-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK3:       omp_offload.failed5:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i32 [[TMP12]], i32 [[TMP14]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK3:       omp_offload.cont6:
-// CHECK3-NEXT:    [[TMP29:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP29]]
+// CHECK3-NEXT:    store i16 [[TMP21]], i16* [[CONV]], align 2
+// CHECK3-NEXT:    [[TMP22:%.*]] = load i32, i32* [[B_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP20]], i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP20]], i32* [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP22]], i32* [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP31:%.*]] = bitcast i8** [[TMP30]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP22]], i32* [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP33]], i8*** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP39]], align 4
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK3-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
+// CHECK3-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK3:       omp_offload.failed6:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i32 [[TMP20]], i32 [[TMP22]]) #[[ATTR3]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK3:       omp_offload.cont7:
+// CHECK3-NEXT:    [[TMP45:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP45]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121

diff --git a/clang/test/OpenMP/target_parallel_num_threads_codegen.cpp b/clang/test/OpenMP/target_parallel_num_threads_codegen.cpp
index cc7e27e..c2c7b1b 100644 (file)
--- a/clang/test/OpenMP/target_parallel_num_threads_codegen.cpp
+++ b/clang/test/OpenMP/target_parallel_num_threads_codegen.cpp
@@ -275,35 +275,69 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK1-NEXT:    [[TMP24:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, i32 3, i8** [[TMP21]], i8** [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP23]])
-// CHECK1-NEXT:    [[TMP25:%.*]] = icmp ne i32 [[TMP24]], 0
-// CHECK1-NEXT:    br i1 [[TMP25]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 1, i32 [[TMP23]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121(%struct.S1* [[THIS1]], i64 [[TMP3]], i64 [[TMP5]]) #[[ATTR3:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to %struct.S1**
-// CHECK1-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP27]], align 8
-// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to double**
-// CHECK1-NEXT:    store double* [[A3]], double** [[TMP29]], align 8
-// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP30]], align 8
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, i32 1, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 1, i32 1024)
-// CHECK1-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK1-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.S1**
+// CHECK1-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to double**
+// CHECK1-NEXT:    store double* [[A3]], double** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 1024, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP35:%.*]] = load double, double* [[A9]], align 8
-// CHECK1-NEXT:    [[CONV10:%.*]] = fptosi double [[TMP35]] to i32
-// CHECK1-NEXT:    ret i32 [[CONV10]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = load double, double* [[A10]], align 8
+// CHECK1-NEXT:    [[CONV11:%.*]] = fptosi double [[TMP51]] to i32
+// CHECK1-NEXT:    ret i32 [[CONV11]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZL7fstatici
@@ -338,41 +372,75 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK1-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP10]])
-// CHECK1-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK1-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP9]], i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 [[TMP10]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP20:%.*]] = icmp ne i32 [[TMP19]], 0
+// CHECK1-NEXT:    br i1 [[TMP20]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104(i64 [[TMP2]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP13:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 32, [[TMP13]]
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 32, [[TMP21]]
 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK1-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK1-NEXT:    [[TMP22:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
 // CHECK1-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED2]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP14]], i32* [[CONV3]], align 4
-// CHECK1-NEXT:    [[TMP15:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED2]], align 8
-// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP15]], i64* [[TMP17]], align 8
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP15]], i64* [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP20]], align 8
-// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK1-NEXT:    [[TMP24:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, i32 1, i8** [[TMP21]], i8** [[TMP22]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP23]])
-// CHECK1-NEXT:    [[TMP25:%.*]] = icmp ne i32 [[TMP24]], 0
-// CHECK1-NEXT:    br i1 [[TMP25]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108(i64 [[TMP15]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[TMP26:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP26]], 1
-// CHECK1-NEXT:    ret i32 [[ADD9]]
+// CHECK1-NEXT:    store i32 [[TMP22]], i32* [[CONV3]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED2]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP23]], i64* [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP23]], i64* [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP29]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP30]], i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 [[TMP31]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK1-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108(i64 [[TMP23]]) #[[ATTR3]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[TMP42:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[ADD10:%.*]] = add nsw i32 [[TMP42]], 1
+// CHECK1-NEXT:    ret i32 [[ADD10]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -390,65 +458,99 @@ int bar(int n){
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 8
 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK1-NEXT:    store i32 0, i32* [[A]], align 4
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 20)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 20, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88() #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i16 1, i16* [[B]], align 2
-// CHECK1-NEXT:    [[TMP2:%.*]] = load i16, i16* [[B]], align 2
-// CHECK1-NEXT:    store i16 [[TMP2]], i16* [[DOTCAPTURE_EXPR_]], align 2
-// CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = load i16, i16* [[B]], align 2
+// CHECK1-NEXT:    store i16 [[TMP10]], i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK1-NEXT:    [[TMP11:%.*]] = load i32, i32* [[A]], align 4
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP3]], i32* [[CONV]], align 4
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i16, i16* [[B]], align 2
+// CHECK1-NEXT:    store i32 [[TMP11]], i32* [[CONV]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = load i16, i16* [[B]], align 2
 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[B_CASTED]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP5]], i16* [[CONV1]], align 2
-// CHECK1-NEXT:    [[TMP6:%.*]] = load i64, i64* [[B_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP7:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK1-NEXT:    store i16 [[TMP13]], i16* [[CONV1]], align 2
+// CHECK1-NEXT:    [[TMP14:%.*]] = load i64, i64* [[B_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
 // CHECK1-NEXT:    [[CONV2:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP7]], i16* [[CONV2]], align 2
-// CHECK1-NEXT:    [[TMP8:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP15]], align 8
-// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP17]], align 8
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP18]], align 8
-// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK1-NEXT:    store i16 [[TMP15]], i16* [[CONV2]], align 2
+// CHECK1-NEXT:    [[TMP16:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP8]], i64* [[TMP20]], align 8
-// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP8]], i64* [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP23]], align 8
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP26:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
-// CHECK1-NEXT:    [[TMP27:%.*]] = zext i16 [[TMP26]] to i32
-// CHECK1-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP27]])
-// CHECK1-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
-// CHECK1-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
-// CHECK1:       omp_offload.failed3:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i64 [[TMP4]], i64 [[TMP6]], i64 [[TMP8]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
-// CHECK1:       omp_offload.cont4:
-// CHECK1-NEXT:    [[TMP30:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP30]]
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP16]], i64* [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP16]], i64* [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
+// CHECK1-NEXT:    store i8* null, i8** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK1-NEXT:    [[TMP35:%.*]] = zext i16 [[TMP34]] to i32
+// CHECK1-NEXT:    [[KERNEL_ARGS3:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP37]], align 4
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP32]], i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP33]], i8*** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 [[TMP35]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]])
+// CHECK1-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK1-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED4:%.*]], label [[OMP_OFFLOAD_CONT5:%.*]]
+// CHECK1:       omp_offload.failed4:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i64 [[TMP12]], i64 [[TMP14]], i64 [[TMP16]]) #[[ATTR3]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT5]]
+// CHECK1:       omp_offload.cont5:
+// CHECK1-NEXT:    [[TMP46:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP46]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121
@@ -730,34 +832,68 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, i32 3, i8** [[TMP21]], i8** [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP23]])
-// CHECK3-NEXT:    [[TMP25:%.*]] = icmp ne i32 [[TMP24]], 0
-// CHECK3-NEXT:    br i1 [[TMP25]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 1, i32 [[TMP23]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121(%struct.S1* [[THIS1]], i32 [[TMP3]], i32 [[TMP5]]) #[[ATTR3:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to %struct.S1**
-// CHECK3-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP27]], align 4
-// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to double**
-// CHECK3-NEXT:    store double* [[A2]], double** [[TMP29]], align 4
-// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP30]], align 4
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, i32 1, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 1, i32 1024)
-// CHECK3-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK3-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK3:       omp_offload.failed6:
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.S1**
+// CHECK3-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to double**
+// CHECK3-NEXT:    store double* [[A2]], double** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 1024, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK3:       omp_offload.failed7:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK3:       omp_offload.cont7:
-// CHECK3-NEXT:    [[A8:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP35:%.*]] = load double, double* [[A8]], align 4
-// CHECK3-NEXT:    [[CONV:%.*]] = fptosi double [[TMP35]] to i32
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK3:       omp_offload.cont8:
+// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = load double, double* [[A9]], align 4
+// CHECK3-NEXT:    [[CONV:%.*]] = fptosi double [[TMP51]] to i32
 // CHECK3-NEXT:    ret i32 [[CONV]]
 //
 //
@@ -792,40 +928,74 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP10]])
-// CHECK3-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK3-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP9]], i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 [[TMP10]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP20:%.*]] = icmp ne i32 [[TMP19]], 0
+// CHECK3-NEXT:    br i1 [[TMP20]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104(i32 [[TMP2]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP13:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 32, [[TMP13]]
+// CHECK3-NEXT:    [[TMP21:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 32, [[TMP21]]
 // CHECK3-NEXT:    store i32 [[ADD]], i32* [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK3-NEXT:    store i32 [[TMP14]], i32* [[DOTCAPTURE_EXPR__CASTED2]], align 4
-// CHECK3-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED2]], align 4
-// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP15]], i32* [[TMP17]], align 4
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP15]], i32* [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP20]], align 4
-// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, i32 1, i8** [[TMP21]], i8** [[TMP22]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP23]])
-// CHECK3-NEXT:    [[TMP25:%.*]] = icmp ne i32 [[TMP24]], 0
-// CHECK3-NEXT:    br i1 [[TMP25]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK3:       omp_offload.failed6:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108(i32 [[TMP15]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK3:       omp_offload.cont7:
-// CHECK3-NEXT:    [[TMP26:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[ADD8:%.*]] = add nsw i32 [[TMP26]], 1
-// CHECK3-NEXT:    ret i32 [[ADD8]]
+// CHECK3-NEXT:    [[TMP22:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK3-NEXT:    store i32 [[TMP22]], i32* [[DOTCAPTURE_EXPR__CASTED2]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED2]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP23]], i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP23]], i32* [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK3-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP29]], i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP30]], i8*** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP39]], align 4
+// CHECK3-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 [[TMP31]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK3-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK3-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108(i32 [[TMP23]]) #[[ATTR3]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK3:       omp_offload.cont8:
+// CHECK3-NEXT:    [[TMP42:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP42]], 1
+// CHECK3-NEXT:    ret i32 [[ADD9]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -843,64 +1013,98 @@ int bar(int n){
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 4
 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK3-NEXT:    store i32 0, i32* [[A]], align 4
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 20)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 20, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88() #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i16 1, i16* [[B]], align 2
-// CHECK3-NEXT:    [[TMP2:%.*]] = load i16, i16* [[B]], align 2
-// CHECK3-NEXT:    store i16 [[TMP2]], i16* [[DOTCAPTURE_EXPR_]], align 2
-// CHECK3-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[A_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP5:%.*]] = load i16, i16* [[B]], align 2
+// CHECK3-NEXT:    [[TMP10:%.*]] = load i16, i16* [[B]], align 2
+// CHECK3-NEXT:    store i16 [[TMP10]], i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK3-NEXT:    [[TMP11:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP11]], i32* [[A_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[A_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = load i16, i16* [[B]], align 2
 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[B_CASTED]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP5]], i16* [[CONV]], align 2
-// CHECK3-NEXT:    [[TMP6:%.*]] = load i32, i32* [[B_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP7:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK3-NEXT:    store i16 [[TMP13]], i16* [[CONV]], align 2
+// CHECK3-NEXT:    [[TMP14:%.*]] = load i32, i32* [[B_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
 // CHECK3-NEXT:    [[CONV1:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR__CASTED]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP7]], i16* [[CONV1]], align 2
-// CHECK3-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP4]], i32* [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP4]], i32* [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP6]], i32* [[TMP15]], align 4
-// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP6]], i32* [[TMP17]], align 4
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP18]], align 4
-// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK3-NEXT:    store i16 [[TMP15]], i16* [[CONV1]], align 2
+// CHECK3-NEXT:    [[TMP16:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP8]], i32* [[TMP20]], align 4
-// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP8]], i32* [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP23]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP26:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
-// CHECK3-NEXT:    [[TMP27:%.*]] = zext i16 [[TMP26]] to i32
-// CHECK3-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 1, i32 [[TMP27]])
-// CHECK3-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
-// CHECK3-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK3:       omp_offload.failed2:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i32 [[TMP4]], i32 [[TMP6]], i32 [[TMP8]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK3:       omp_offload.cont3:
-// CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP30]]
+// CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP14]], i32* [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP14]], i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP16]], i32* [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP16]], i32* [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8* null, i8** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK3-NEXT:    [[TMP35:%.*]] = zext i16 [[TMP34]] to i32
+// CHECK3-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP32]], i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP33]], i8*** [[TMP39]], align 4
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1, i32 [[TMP35]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK3-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK3-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK3:       omp_offload.failed3:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i32 [[TMP12]], i32 [[TMP14]], i32 [[TMP16]]) #[[ATTR3]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK3:       omp_offload.cont4:
+// CHECK3-NEXT:    [[TMP46:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP46]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121

diff --git a/clang/test/OpenMP/target_parallel_uses_allocators_codegen.cpp b/clang/test/OpenMP/target_parallel_uses_allocators_codegen.cpp
index 8294185..d4bed46 100644 (file)
--- a/clang/test/OpenMP/target_parallel_uses_allocators_codegen.cpp
+++ b/clang/test/OpenMP/target_parallel_uses_allocators_codegen.cpp
@@ -65,7 +65,7 @@ void foo() {
   omp_alloctrait_t traits[10];
   omp_allocator_handle_t my_allocator;
 
-// CHECK: [[RES:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @.[[TGT_REGION:.+]].region_id, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZES]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPTYPES]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
+// CHECK: [[RES:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i32 0, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
 // CHECK: [[CMP:%.+]] = icmp ne i32 [[RES]], 0
 // CHECK: br i1 [[CMP]], label %[[FAILED:.+]], label %[[DONE:.+]]
 // CHECK: [[FAILED]]:

diff --git a/clang/test/OpenMP/target_simd_codegen.cpp b/clang/test/OpenMP/target_simd_codegen.cpp
index 9e73a53..8931397 100644 (file)
--- a/clang/test/OpenMP/target_simd_codegen.cpp
+++ b/clang/test/OpenMP/target_simd_codegen.cpp
@@ -147,7 +147,11 @@ int foo(int n) {
     a += 1;
   }
 
-  // CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 3, i8** [[BP:%[^,]+]], i8** [[P:%[^,]+]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* [[SIZET2]], i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* [[MAPT2]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 1)
+  // CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 1, i32 1, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CHECK-DAG:   [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CHECK-DAG:   store i8** [[BP:%.+]], i8*** [[BPARG]]
+  // CHECK-DAG:   [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CHECK-DAG:   store i8** [[P:%.+]], i8*** [[PARG]]
   // CHECK-DAG:   [[BP]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[BPR:%[^,]+]], i32 0, i32 0
   // CHECK-DAG:   [[P]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[PR:%[^,]+]], i32 0, i32 0
   // CHECK-DAG:   [[BPADDR0:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[BPR]], i32 0, i32 0
@@ -184,7 +188,11 @@ int foo(int n) {
   // CHECK:       [[IF:%.+]] = icmp sgt i32 {{[^,]+}}, 10
   // CHECK:       br i1 [[IF]], label %[[IFTHEN:[^,]+]], label %[[IFELSE:[^,]+]]
   // CHECK:       [[IFTHEN]]
-  // CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 2, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[SIZET3]], i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT3]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 1)
+  // CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 1, i32 1, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+  // CHECK-DAG:   [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+  // CHECK-DAG:   store i8** [[BPR:%.+]], i8*** [[BPARG]]
+  // CHECK-DAG:   [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+  // CHECK-DAG:   store i8** [[PR:%.+]], i8*** [[PARG]]
   // CHECK-DAG:   [[BPR]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[BP:%[^,]+]], i32 0, i32 0
   // CHECK-DAG:   [[PR]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[P:%[^,]+]], i32 0, i32 0
 
@@ -240,89 +248,96 @@ int foo(int n) {
   // CHECK-32:    [[CNSZSIZE:%.+]] = mul nuw i32 [[CNELEMSIZE2]], 8
   // CHECK-32:    [[CNSIZE:%.+]] = sext i32 [[CNSZSIZE]] to i64
 
-  // CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 9, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SR:%[^,]+]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* [[MAPT4]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 1)
-  // CHECK-DAG:   [[BPR]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP:%[^,]+]], i32 0, i32 0
-  // CHECK-DAG:   [[PR]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P:%[^,]+]], i32 0, i32 0
-  // CHECK-DAG:   [[SR]] = getelementptr inbounds [9 x i64], [9 x i64]* [[S:%[^,]+]], i32 0, i32 0
-
-  // CHECK-DAG:   [[BPADDR0:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX0:[0-9]+]]
-  // CHECK-DAG:   [[PADDR0:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX0]]
-  // CHECK-DAG:   [[BPADDR1:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX1:[0-9]+]]
-  // CHECK-DAG:   [[PADDR1:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX1]]
-  // CHECK-DAG:   [[BPADDR2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX2:[0-9]+]]
-  // CHECK-DAG:   [[PADDR2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX2]]
-  // CHECK-DAG:   [[SADDR3:%.+]] = getelementptr inbounds [9 x i64], [9 x i64]* [[S]], i32 0, i32 [[IDX3:[0-9]+]]
-  // CHECK-DAG:   [[BPADDR3:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX3]]
-  // CHECK-DAG:   [[PADDR3:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX3]]
-  // CHECK-DAG:   [[BPADDR4:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX4:[0-9]+]]
-  // CHECK-DAG:   [[PADDR4:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX4]]
-  // CHECK-DAG:   [[BPADDR5:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX5:[0-9]+]]
-  // CHECK-DAG:   [[PADDR5:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX5]]
-  // CHECK-DAG:   [[BPADDR6:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX6:[0-9]+]]
-  // CHECK-DAG:   [[PADDR6:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX6]]
-  // CHECK-DAG:   [[SADDR7:%.+]] = getelementptr inbounds [9 x i64], [9 x i64]* [[S]], i32 0, i32 [[IDX7:[0-9]+]]
-  // CHECK-DAG:   [[BPADDR7:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX7]]
-  // CHECK-DAG:   [[PADDR7:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX7]]
-  // CHECK-DAG:   [[BPADDR8:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX8:[0-9]+]]
-  // CHECK-DAG:   [[PADDR8:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX8]]
-
-  // The names below are not necessarily consistent with the names used for the
-  // addresses above as some are repeated.
-  // CHECK-DAG:   store i[[SZ]] [[A_CVAL]], i[[SZ]]* [[CBPADDR0:%.+]],
-  // CHECK-DAG:   store i[[SZ]] [[A_CVAL]], i[[SZ]]* [[CPADDR0:%.+]],
-  // CHECK-DAG:   [[CBPADDR0]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
-  // CHECK-DAG:   [[CPADDR0]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
-
-  // CHECK-DAG:   store [10 x float]* %{{.+}}, [10 x float]** [[CBPADDR1:%.+]],
-  // CHECK-DAG:   store [10 x float]* %{{.+}}, [10 x float]** [[CPADDR1:%.+]],
-  // CHECK-DAG:   [[CBPADDR1]] = bitcast i8** {{%[^,]+}} to [10 x float]**
-  // CHECK-DAG:   [[CPADDR1]] = bitcast i8** {{%[^,]+}} to [10 x float]**
-
-  // CHECK-DAG:   store i[[SZ]] %{{.+}}, i[[SZ]]* [[CBPADDR2:%.+]],
-  // CHECK-DAG:   store i[[SZ]] %{{.+}}, i[[SZ]]* [[CPADDR2:%.+]],
-  // CHECK-DAG:   [[CBPADDR2]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
-  // CHECK-DAG:   [[CPADDR2]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
-
-  // CHECK-DAG:   store float* %{{.+}}, float** [[CBPADDR3:%.+]],
-  // CHECK-DAG:   store float* %{{.+}}, float** [[CPADDR3:%.+]],
-  // CHECK-DAG:   [[CBPADDR3]] = bitcast i8** {{%[^,]+}} to float**
-  // CHECK-DAG:   [[CPADDR3]] = bitcast i8** {{%[^,]+}} to float**
-  // CHECK-DAG:   store i64 [[BNSIZE]], i64* {{%[^,]+}}
-
-  // CHECK-DAG:   store [5 x [10 x double]]* %{{.+}}, [5 x [10 x double]]** [[CBPADDR4:%.+]],
-  // CHECK-DAG:   store [5 x [10 x double]]* %{{.+}}, [5 x [10 x double]]** [[CPADDR4:%.+]],
-  // CHECK-DAG:   [[CBPADDR4]] = bitcast i8** {{%[^,]+}} to [5 x [10 x double]]**
-  // CHECK-DAG:   [[CPADDR4]] = bitcast i8** {{%[^,]+}} to [5 x [10 x double]]**
-
-  // CHECK-DAG:   store i[[SZ]] 5, i[[SZ]]* [[CBPADDR5:%.+]],
-  // CHECK-DAG:   store i[[SZ]] 5, i[[SZ]]* [[CPADDR5:%.+]],
-  // CHECK-DAG:   [[CBPADDR5]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
-  // CHECK-DAG:   [[CPADDR5]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
-
-  // CHECK-DAG:   store i[[SZ]] %{{.+}}, i[[SZ]]* [[CBPADDR6:%.+]],
-  // CHECK-DAG:   store i[[SZ]] %{{.+}}, i[[SZ]]* [[CPADDR6:%.+]],
-  // CHECK-DAG:   [[CBPADDR6]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
-  // CHECK-DAG:   [[CPADDR6]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
-
-  // CHECK-DAG:   store double* %{{.+}}, double** [[CBPADDR7:%.+]],
-  // CHECK-DAG:   store double* %{{.+}}, double** [[CPADDR7:%.+]],
-  // CHECK-DAG:   [[CBPADDR7]] = bitcast i8** {{%[^,]+}} to double**
-  // CHECK-DAG:   [[CPADDR7]] = bitcast i8** {{%[^,]+}} to double**
-  // CHECK-DAG:   store i64 [[CNSIZE]], i64* {{%[^,]+}}
-
-  // CHECK-DAG:   store [[TT]]* %{{.+}}, [[TT]]** [[CBPADDR8:%.+]],
-  // CHECK-DAG:   store [[TT]]* %{{.+}}, [[TT]]** [[CPADDR8:%.+]],
-  // CHECK-DAG:   [[CBPADDR8]] = bitcast i8** {{%[^,]+}} to [[TT]]**
-  // CHECK-DAG:   [[CPADDR8]] = bitcast i8** {{%[^,]+}} to [[TT]]**
+// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 1, i32 1, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CHECK-DAG:   [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CHECK-DAG:   store i8** [[BPR:%.+]], i8*** [[BPARG]]
+// CHECK-DAG:   [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CHECK-DAG:   store i8** [[PR:%.+]], i8*** [[PARG]]
+// CHECK-DAG:   [[SARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 4
+// CHECK-DAG:   store i64* [[SR:%.+]], i64** [[SARG]]
+// CHECK-DAG:   [[BPR]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP:%[^,]+]], i32 0, i32 0
+// CHECK-DAG:   [[PR]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P:%[^,]+]], i32 0, i32 0
+// CHECK-DAG:   [[SR]] = getelementptr inbounds [9 x i64], [9 x i64]* [[S:%[^,]+]], i32 0, i32 0
+
+// CHECK-DAG:   [[BPADDR0:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX0:[0-9]+]]
+// CHECK-DAG:   [[PADDR0:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX0]]
+// CHECK-DAG:   [[BPADDR1:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX1:[0-9]+]]
+// CHECK-DAG:   [[PADDR1:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX1]]
+// CHECK-DAG:   [[BPADDR2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX2:[0-9]+]]
+// CHECK-DAG:   [[PADDR2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX2]]
+// CHECK-DAG:   [[SADDR3:%.+]] = getelementptr inbounds [9 x i64], [9 x i64]* [[S]], i32 0, i32 [[IDX3:[0-9]+]]
+// CHECK-DAG:   [[BPADDR3:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX3]]
+// CHECK-DAG:   [[PADDR3:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX3]]
+// CHECK-DAG:   [[BPADDR4:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX4:[0-9]+]]
+// CHECK-DAG:   [[PADDR4:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX4]]
+// CHECK-DAG:   [[BPADDR5:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX5:[0-9]+]]
+// CHECK-DAG:   [[PADDR5:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX5]]
+// CHECK-DAG:   [[BPADDR6:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX6:[0-9]+]]
+// CHECK-DAG:   [[PADDR6:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX6]]
+// CHECK-DAG:   [[SADDR7:%.+]] = getelementptr inbounds [9 x i64], [9 x i64]* [[S]], i32 0, i32 [[IDX7:[0-9]+]]
+// CHECK-DAG:   [[BPADDR7:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX7]]
+// CHECK-DAG:   [[PADDR7:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX7]]
+// CHECK-DAG:   [[BPADDR8:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BP]], i32 0, i32 [[IDX8:[0-9]+]]
+// CHECK-DAG:   [[PADDR8:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[P]], i32 0, i32 [[IDX8]]
 
-  // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
-  // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
+// The names below are not necessarily consistent with the names used for the
+// addresses above as some are repeated.
+// CHECK-DAG:   store i[[SZ]] [[A_CVAL]], i[[SZ]]* [[CBPADDR0:%.+]],
+// CHECK-DAG:   store i[[SZ]] [[A_CVAL]], i[[SZ]]* [[CPADDR0:%.+]],
+// CHECK-DAG:   [[CBPADDR0]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
+// CHECK-DAG:   [[CPADDR0]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
+
+// CHECK-DAG:   store [10 x float]* %{{.+}}, [10 x float]** [[CBPADDR1:%.+]],
+// CHECK-DAG:   store [10 x float]* %{{.+}}, [10 x float]** [[CPADDR1:%.+]],
+// CHECK-DAG:   [[CBPADDR1]] = bitcast i8** {{%[^,]+}} to [10 x float]**
+// CHECK-DAG:   [[CPADDR1]] = bitcast i8** {{%[^,]+}} to [10 x float]**
+
+// CHECK-DAG:   store i[[SZ]] %{{.+}}, i[[SZ]]* [[CBPADDR2:%.+]],
+// CHECK-DAG:   store i[[SZ]] %{{.+}}, i[[SZ]]* [[CPADDR2:%.+]],
+// CHECK-DAG:   [[CBPADDR2]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
+// CHECK-DAG:   [[CPADDR2]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
 
-  // CHECK:       [[FAIL]]
-  // CHECK:       call void [[HVT4:@.+]]({{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}})
-  // CHECK-NEXT:  br label %[[END]]
-  // CHECK:       [[END]]
-  #pragma omp target simd if(target: n>20)
+// CHECK-DAG:   store float* %{{.+}}, float** [[CBPADDR3:%.+]],
+// CHECK-DAG:   store float* %{{.+}}, float** [[CPADDR3:%.+]],
+// CHECK-DAG:   [[CBPADDR3]] = bitcast i8** {{%[^,]+}} to float**
+// CHECK-DAG:   [[CPADDR3]] = bitcast i8** {{%[^,]+}} to float**
+// CHECK-DAG:   store i64 [[BNSIZE]], i64* {{%[^,]+}}
+
+// CHECK-DAG:   store [5 x [10 x double]]* %{{.+}}, [5 x [10 x double]]** [[CBPADDR4:%.+]],
+// CHECK-DAG:   store [5 x [10 x double]]* %{{.+}}, [5 x [10 x double]]** [[CPADDR4:%.+]],
+// CHECK-DAG:   [[CBPADDR4]] = bitcast i8** {{%[^,]+}} to [5 x [10 x double]]**
+// CHECK-DAG:   [[CPADDR4]] = bitcast i8** {{%[^,]+}} to [5 x [10 x double]]**
+
+// CHECK-DAG:   store i[[SZ]] 5, i[[SZ]]* [[CBPADDR5:%.+]],
+// CHECK-DAG:   store i[[SZ]] 5, i[[SZ]]* [[CPADDR5:%.+]],
+// CHECK-DAG:   [[CBPADDR5]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
+// CHECK-DAG:   [[CPADDR5]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
+
+// CHECK-DAG:   store i[[SZ]] %{{.+}}, i[[SZ]]* [[CBPADDR6:%.+]],
+// CHECK-DAG:   store i[[SZ]] %{{.+}}, i[[SZ]]* [[CPADDR6:%.+]],
+// CHECK-DAG:   [[CBPADDR6]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
+// CHECK-DAG:   [[CPADDR6]] = bitcast i8** {{%[^,]+}} to i[[SZ]]*
+
+// CHECK-DAG:   store double* %{{.+}}, double** [[CBPADDR7:%.+]],
+// CHECK-DAG:   store double* %{{.+}}, double** [[CPADDR7:%.+]],
+// CHECK-DAG:   [[CBPADDR7]] = bitcast i8** {{%[^,]+}} to double**
+// CHECK-DAG:   [[CPADDR7]] = bitcast i8** {{%[^,]+}} to double**
+// CHECK-DAG:   store i64 [[CNSIZE]], i64* {{%[^,]+}}
+
+// CHECK-DAG:   store [[TT]]* %{{.+}}, [[TT]]** [[CBPADDR8:%.+]],
+// CHECK-DAG:   store [[TT]]* %{{.+}}, [[TT]]** [[CPADDR8:%.+]],
+// CHECK-DAG:   [[CBPADDR8]] = bitcast i8** {{%[^,]+}} to [[TT]]**
+// CHECK-DAG:   [[CPADDR8]] = bitcast i8** {{%[^,]+}} to [[TT]]**
+
+// CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
+// CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
+
+// CHECK:       [[FAIL]]
+// CHECK:       call void [[HVT4:@.+]]({{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}}, {{[^,]+}})
+// CHECK-NEXT:  br label %[[END]]
+// CHECK:       [[END]]
+#pragma omp target simd if (target \
+                            : n > 20)
   for (unsigned char it = 'z'; it >= 'a'; it+=-1) {
     a += 1;
     b[2] += 1.0;
@@ -344,9 +359,8 @@ int foo(int n) {
 // CHECK:       ret void
 // CHECK-NEXT:  }
 
-
 // CHECK:       define internal {{.*}}i32 [[OMP_TASK_ENTRY]](i32 {{.*}}%0, [[KMP_TASK_T_WITH_PRIVATES]]* noalias noundef %1)
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 1, i32 0, i8* null, i32 0, i8* null)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 1, i32 1, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]], i32 0, i8* null, i32 0, i8* null)
 // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // CHECK:       [[FAIL]]
@@ -354,7 +368,6 @@ int foo(int n) {
 // CHECK-NEXT:  br label %[[END]]
 // CHECK:       [[END]]
 
-
 // CHECK:       define internal void [[HVT1]](i[[SZ]] noundef %{{.+}}, i{{32|64}}{{[*]*.*}} %{{.+}})
 // CHECK:       [[AA_ADDR:%.+]] = alloca i[[SZ]], align
 // CHECK:       store i[[SZ]] %{{.+}}, i[[SZ]]* [[AA_ADDR]], align
@@ -521,8 +534,11 @@ int bar(int n){
 // CHECK-32:    [[CSZSIZE:%.+]] = mul nuw i32 [[CELEMSIZE2]], 2
 // CHECK-32:    [[CSIZE:%.+]] = sext i32 [[CSZSIZE]] to i64
 
-// OMP45-DAG:   [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 5, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SR:%[^,]+]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* [[MAPT7]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 1)
-// OMP50-DAG:   [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 6, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SR:%[^,]+]], i64* getelementptr inbounds ([6 x  i64], [6 x  i64]* [[MAPT7]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 1)
+// OMP45-DAG:   [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 1, i32 1, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// OMP45-DAG:   [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// OMP45-DAG:   store i8** [[BPR:%.+]], i8*** [[BPARG]]
+// OMP45-DAG:   [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// OMP45-DAG:   store i8** [[PR:%.+]], i8*** [[PARG]]
 // OMP45-DAG:   [[BPR]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[BP:%.+]], i32 0, i32 0
 // OMP45-DAG:   [[PR]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[P:%.+]], i32 0, i32 0
 // OMP45-DAG:   [[BPADDR0:%.+]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[BP]], i32 [[IDX0:[0-9]+]]
@@ -536,6 +552,11 @@ int bar(int n){
 // OMP45-DAG:   [[SADDR4:%.+]] = getelementptr inbounds [5 x i64], [5 x i64]* [[S]], i32 [[IDX4:[0-9]+]]
 // OMP45-DAG:   [[BPADDR4:%.+]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[BP]], i32 [[IDX4]]
 // OMP45-DAG:   [[PADDR4:%.+]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[P]], i32 [[IDX4]]
+// OMP50-DAG:   [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 1, i32 1, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// OMP50-DAG:   [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// OMP50-DAG:   store i8** [[BPR:%.+]], i8*** [[BPARG]]
+// OMP50-DAG:   [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// OMP50-DAG:   store i8** [[PR:%.+]], i8*** [[PARG]]
 // OMP50-DAG:   [[BPR]] = getelementptr inbounds [6 x  i8*], [6 x  i8*]* [[BP:%.+]], i32 0, i32 0
 // OMP50-DAG:   [[PR]] = getelementptr inbounds [6 x  i8*], [6 x  i8*]* [[P:%.+]], i32 0, i32 0
 // OMP50-DAG:   [[BPADDR0:%.+]] = getelementptr inbounds [6 x  i8*], [6 x  i8*]* [[BP]], i32 [[IDX0:[0-9]+]]
@@ -600,7 +621,11 @@ int bar(int n){
 // CHECK:       [[IF:%.+]] = icmp sgt i32 {{[^,]+}}, 50
 // CHECK:       br i1 [[IF]], label %[[IFTHEN:[^,]+]], label %[[IFELSE:[^,]+]]
 // CHECK:       [[IFTHEN]]
-// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 4, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* [[SIZET6]], i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* [[MAPT6]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 1)
+// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 1, i32 1, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CHECK-DAG:   [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CHECK-DAG:   store i8** [[BPR:%.+]], i8*** [[BPARG]]
+// CHECK-DAG:   [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CHECK-DAG:   store i8** [[PR:%.+]], i8*** [[PARG]]
 // CHECK-DAG:   [[BPR]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[BP:%.+]], i32 0, i32 0
 // CHECK-DAG:   [[PR]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[P:%.+]], i32 0, i32 0
 
@@ -650,7 +675,11 @@ int bar(int n){
 // CHECK:       [[IF:%.+]] = icmp sgt i32 {{[^,]+}}, 40
 // CHECK:       br i1 [[IF]], label %[[IFTHEN:[^,]+]], label %[[IFELSE:[^,]+]]
 // CHECK:       [[IFTHEN]]
-// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @{{[^,]+}}, i32 3, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* [[SIZET5]], i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* [[MAPT5]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 1)
+// CHECK-DAG:   [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE:.+]], i32 1, i32 1, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* [[ARGS:%.+]])
+// CHECK-DAG:   [[BPARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 2
+// CHECK-DAG:   store i8** [[BPR:%.+]], i8*** [[BPARG]]
+// CHECK-DAG:   [[PARG:%.+]] = getelementptr inbounds {{.+}}[[ARGS]], i32 0, i32 3
+// CHECK-DAG:   store i8** [[PR:%.+]], i8*** [[PARG]]
 // CHECK-DAG:   [[BPR]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[BP:%.+]], i32 0, i32 0
 // CHECK-DAG:   [[PR]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[P:%.+]], i32 0, i32 0
 

diff --git a/clang/test/OpenMP/target_simd_depend_codegen.cpp b/clang/test/OpenMP/target_simd_depend_codegen.cpp
index 55d6e1f..aad4f48 100644 (file)
--- a/clang/test/OpenMP/target_simd_depend_codegen.cpp
+++ b/clang/test/OpenMP/target_simd_depend_codegen.cpp
@@ -186,7 +186,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 0
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 1, i32 1)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 1, i32 1, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // CHECK:       [[FAIL]]
@@ -204,7 +204,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 2
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 2, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SZT]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 1, i32 0, i8* null, i32 0, i8* null)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 1, i32 1, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 
 // CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]

diff --git a/clang/test/OpenMP/target_simd_uses_allocators_codegen.cpp b/clang/test/OpenMP/target_simd_uses_allocators_codegen.cpp
index 522abef..940de58 100644 (file)
--- a/clang/test/OpenMP/target_simd_uses_allocators_codegen.cpp
+++ b/clang/test/OpenMP/target_simd_uses_allocators_codegen.cpp
@@ -65,7 +65,7 @@ void foo() {
   omp_alloctrait_t traits[10];
   omp_allocator_handle_t my_allocator;
 
-// CHECK: [[RES:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @.[[TGT_REGION:.+]].region_id, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZES]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPTYPES]], i32 0, i32 0), i8** null, i8** null, i32 1, i32 1)
+// CHECK: [[RES:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 1, i32 1, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
 // CHECK: [[CMP:%.+]] = icmp ne i32 [[RES]], 0
 // CHECK: br i1 [[CMP]], label %[[FAILED:.+]], label %[[DONE:.+]]
 // CHECK: [[FAILED]]:

diff --git a/clang/test/OpenMP/target_teams_codegen.cpp b/clang/test/OpenMP/target_teams_codegen.cpp
index 76946df..bc51679 100644 (file)
--- a/clang/test/OpenMP/target_teams_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_codegen.cpp
@@ -333,20 +333,20 @@ int bar(int n){
 // CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS16:%.*]] = alloca [2 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS17:%.*]] = alloca [2 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS18:%.*]] = alloca [2 x i8*], align 8
-// CHECK1-NEXT:    [[A_CASTED21:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS25:%.*]] = alloca [9 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS26:%.*]] = alloca [9 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS27:%.*]] = alloca [9 x i8*], align 8
+// CHECK1-NEXT:    [[A_CASTED22:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS26:%.*]] = alloca [9 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS27:%.*]] = alloca [9 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS28:%.*]] = alloca [9 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 8
 // CHECK1-NEXT:    [[NN:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[NN_CASTED:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[NN_CASTED38:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS40:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS41:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS42:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS35:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS36:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS37:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[NN_CASTED41:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS43:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS44:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS45:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK1-NEXT:    store i32 0, i32* [[A]], align 4
@@ -456,206 +456,291 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP65]], align 8
 // CHECK1-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP68:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP66]], i8** [[TMP67]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP69:%.*]] = icmp ne i32 [[TMP68]], 0
-// CHECK1-NEXT:    br i1 [[TMP69]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK1-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP69]], align 4
+// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP66]], i8*** [[TMP70]], align 8
+// CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP67]], i8*** [[TMP71]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 8
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP75]], align 8
+// CHECK1-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK1-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i64 [[TMP60]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP70:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP78:%.*]] = load i32, i32* [[A]], align 4
 // CHECK1-NEXT:    [[CONV13:%.*]] = bitcast i64* [[A_CASTED12]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP70]], i32* [[CONV13]], align 4
-// CHECK1-NEXT:    [[TMP71:%.*]] = load i64, i64* [[A_CASTED12]], align 8
-// CHECK1-NEXT:    [[TMP72:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK1-NEXT:    store i32 [[TMP78]], i32* [[CONV13]], align 4
+// CHECK1-NEXT:    [[TMP79:%.*]] = load i64, i64* [[A_CASTED12]], align 8
+// CHECK1-NEXT:    [[TMP80:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK1-NEXT:    [[CONV15:%.*]] = bitcast i64* [[AA_CASTED14]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP72]], i16* [[CONV15]], align 2
-// CHECK1-NEXT:    [[TMP73:%.*]] = load i64, i64* [[AA_CASTED14]], align 8
-// CHECK1-NEXT:    [[TMP74:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP74]], 10
+// CHECK1-NEXT:    store i16 [[TMP80]], i16* [[CONV15]], align 2
+// CHECK1-NEXT:    [[TMP81:%.*]] = load i64, i64* [[AA_CASTED14]], align 8
+// CHECK1-NEXT:    [[TMP82:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP82]], 10
 // CHECK1-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK1:       omp_if.then:
-// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP71]], i64* [[TMP76]], align 8
-// CHECK1-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP71]], i64* [[TMP78]], align 8
-// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP79]], align 8
-// CHECK1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP73]], i64* [[TMP81]], align 8
-// CHECK1-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP73]], i64* [[TMP83]], align 8
-// CHECK1-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP84]], align 8
-// CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP87:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP85]], i8** [[TMP86]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP88:%.*]] = icmp ne i32 [[TMP87]], 0
-// CHECK1-NEXT:    br i1 [[TMP88]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
-// CHECK1:       omp_offload.failed19:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
-// CHECK1:       omp_offload.cont20:
+// CHECK1-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP79]], i64* [[TMP84]], align 8
+// CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP79]], i64* [[TMP86]], align 8
+// CHECK1-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP87]], align 8
+// CHECK1-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP81]], i64* [[TMP89]], align 8
+// CHECK1-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP81]], i64* [[TMP91]], align 8
+// CHECK1-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP92]], align 8
+// CHECK1-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP95]], align 4
+// CHECK1-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP96]], align 4
+// CHECK1-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP93]], i8*** [[TMP97]], align 8
+// CHECK1-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP94]], i8*** [[TMP98]], align 8
+// CHECK1-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP99]], align 8
+// CHECK1-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP100]], align 8
+// CHECK1-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP101]], align 8
+// CHECK1-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP102]], align 8
+// CHECK1-NEXT:    [[TMP103:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK1-NEXT:    [[TMP104:%.*]] = icmp ne i32 [[TMP103]], 0
+// CHECK1-NEXT:    br i1 [[TMP104]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK1:       omp_offload.failed20:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP79]], i64 [[TMP81]]) #[[ATTR3]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK1:       omp_offload.cont21:
 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK1:       omp_if.else:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR3]]
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP79]], i64 [[TMP81]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP89:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    [[CONV22:%.*]] = bitcast i64* [[A_CASTED21]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP89]], i32* [[CONV22]], align 4
-// CHECK1-NEXT:    [[TMP90:%.*]] = load i64, i64* [[A_CASTED21]], align 8
-// CHECK1-NEXT:    [[TMP91:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP23:%.*]] = icmp sgt i32 [[TMP91]], 20
-// CHECK1-NEXT:    br i1 [[CMP23]], label [[OMP_IF_THEN24:%.*]], label [[OMP_IF_ELSE30:%.*]]
-// CHECK1:       omp_if.then24:
-// CHECK1-NEXT:    [[TMP92:%.*]] = mul nuw i64 [[TMP2]], 4
-// CHECK1-NEXT:    [[TMP93:%.*]] = mul nuw i64 5, [[TMP5]]
-// CHECK1-NEXT:    [[TMP94:%.*]] = mul nuw i64 [[TMP93]], 8
-// CHECK1-NEXT:    [[TMP95:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP95]], i8* align 8 bitcast ([9 x i64]* @.offload_sizes.10 to i8*), i64 72, i1 false)
-// CHECK1-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP90]], i64* [[TMP97]], align 8
-// CHECK1-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP90]], i64* [[TMP99]], align 8
-// CHECK1-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP100]], align 8
-// CHECK1-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
-// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 8
-// CHECK1-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
-// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 8
-// CHECK1-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP105]], align 8
-// CHECK1-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP107]], align 8
-// CHECK1-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP109]], align 8
-// CHECK1-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP110]], align 8
-// CHECK1-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to float**
-// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP112]], align 8
-// CHECK1-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
-// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP114]], align 8
-// CHECK1-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK1-NEXT:    store i64 [[TMP92]], i64* [[TMP115]], align 8
-// CHECK1-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 3
+// CHECK1-NEXT:    [[TMP105:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[CONV23:%.*]] = bitcast i64* [[A_CASTED22]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP105]], i32* [[CONV23]], align 4
+// CHECK1-NEXT:    [[TMP106:%.*]] = load i64, i64* [[A_CASTED22]], align 8
+// CHECK1-NEXT:    [[TMP107:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP24:%.*]] = icmp sgt i32 [[TMP107]], 20
+// CHECK1-NEXT:    br i1 [[CMP24]], label [[OMP_IF_THEN25:%.*]], label [[OMP_IF_ELSE32:%.*]]
+// CHECK1:       omp_if.then25:
+// CHECK1-NEXT:    [[TMP108:%.*]] = mul nuw i64 [[TMP2]], 4
+// CHECK1-NEXT:    [[TMP109:%.*]] = mul nuw i64 5, [[TMP5]]
+// CHECK1-NEXT:    [[TMP110:%.*]] = mul nuw i64 [[TMP109]], 8
+// CHECK1-NEXT:    [[TMP111:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP111]], i8* align 8 bitcast ([9 x i64]* @.offload_sizes.10 to i8*), i64 72, i1 false)
+// CHECK1-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP106]], i64* [[TMP113]], align 8
+// CHECK1-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP106]], i64* [[TMP115]], align 8
+// CHECK1-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 0
 // CHECK1-NEXT:    store i8* null, i8** [[TMP116]], align 8
-// CHECK1-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to [5 x [10 x double]]**
-// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP118]], align 8
-// CHECK1-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
-// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 8
-// CHECK1-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 4
+// CHECK1-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to [10 x float]**
+// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP118]], align 8
+// CHECK1-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [10 x float]**
+// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP120]], align 8
+// CHECK1-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 1
 // CHECK1-NEXT:    store i8* null, i8** [[TMP121]], align 8
-// CHECK1-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP123:%.*]] = bitcast i8** [[TMP122]] to i64*
-// CHECK1-NEXT:    store i64 5, i64* [[TMP123]], align 8
-// CHECK1-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 5
+// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP123]], align 8
+// CHECK1-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP125:%.*]] = bitcast i8** [[TMP124]] to i64*
-// CHECK1-NEXT:    store i64 5, i64* [[TMP125]], align 8
-// CHECK1-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 5
+// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP125]], align 8
+// CHECK1-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 2
 // CHECK1-NEXT:    store i8* null, i8** [[TMP126]], align 8
-// CHECK1-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP128]], align 8
-// CHECK1-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP130]], align 8
-// CHECK1-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 6
-// CHECK1-NEXT:    store i8* null, i8** [[TMP131]], align 8
-// CHECK1-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to double**
-// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP133]], align 8
-// CHECK1-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to double**
-// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP135]], align 8
-// CHECK1-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK1-NEXT:    store i64 [[TMP94]], i64* [[TMP136]], align 8
-// CHECK1-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 7
+// CHECK1-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to float**
+// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP128]], align 8
+// CHECK1-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to float**
+// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP130]], align 8
+// CHECK1-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK1-NEXT:    store i64 [[TMP108]], i64* [[TMP131]], align 8
+// CHECK1-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 3
+// CHECK1-NEXT:    store i8* null, i8** [[TMP132]], align 8
+// CHECK1-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to [5 x [10 x double]]**
+// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP134]], align 8
+// CHECK1-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to [5 x [10 x double]]**
+// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP136]], align 8
+// CHECK1-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 4
 // CHECK1-NEXT:    store i8* null, i8** [[TMP137]], align 8
-// CHECK1-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to %struct.TT**
-// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP139]], align 8
-// CHECK1-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to %struct.TT**
-// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP141]], align 8
-// CHECK1-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 8
+// CHECK1-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i64*
+// CHECK1-NEXT:    store i64 5, i64* [[TMP139]], align 8
+// CHECK1-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i64*
+// CHECK1-NEXT:    store i64 5, i64* [[TMP141]], align 8
+// CHECK1-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 5
 // CHECK1-NEXT:    store i8* null, i8** [[TMP142]], align 8
-// CHECK1-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP146:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP143]], i8** [[TMP144]], i64* [[TMP145]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.11, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP147:%.*]] = icmp ne i32 [[TMP146]], 0
-// CHECK1-NEXT:    br i1 [[TMP147]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK1:       omp_offload.failed28:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK1:       omp_offload.cont29:
-// CHECK1-NEXT:    br label [[OMP_IF_END31:%.*]]
-// CHECK1:       omp_if.else30:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_IF_END31]]
-// CHECK1:       omp_if.end31:
+// CHECK1-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 6
+// CHECK1-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP144]], align 8
+// CHECK1-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 6
+// CHECK1-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP146]], align 8
+// CHECK1-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 6
+// CHECK1-NEXT:    store i8* null, i8** [[TMP147]], align 8
+// CHECK1-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to double**
+// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP149]], align 8
+// CHECK1-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to double**
+// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP151]], align 8
+// CHECK1-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK1-NEXT:    store i64 [[TMP110]], i64* [[TMP152]], align 8
+// CHECK1-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 7
+// CHECK1-NEXT:    store i8* null, i8** [[TMP153]], align 8
+// CHECK1-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP155:%.*]] = bitcast i8** [[TMP154]] to %struct.TT**
+// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP155]], align 8
+// CHECK1-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to %struct.TT**
+// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP157]], align 8
+// CHECK1-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 8
+// CHECK1-NEXT:    store i8* null, i8** [[TMP158]], align 8
+// CHECK1-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS29:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP162]], align 4
+// CHECK1-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 9, i32* [[TMP163]], align 4
+// CHECK1-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP159]], i8*** [[TMP164]], align 8
+// CHECK1-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP160]], i8*** [[TMP165]], align 8
+// CHECK1-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP161]], i64** [[TMP166]], align 8
+// CHECK1-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.11, i32 0, i32 0), i64** [[TMP167]], align 8
+// CHECK1-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP168]], align 8
+// CHECK1-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP169]], align 8
+// CHECK1-NEXT:    [[TMP170:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]])
+// CHECK1-NEXT:    [[TMP171:%.*]] = icmp ne i32 [[TMP170]], 0
+// CHECK1-NEXT:    br i1 [[TMP171]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
+// CHECK1:       omp_offload.failed30:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP106]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
+// CHECK1:       omp_offload.cont31:
+// CHECK1-NEXT:    br label [[OMP_IF_END33:%.*]]
+// CHECK1:       omp_if.else32:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP106]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
+// CHECK1-NEXT:    br label [[OMP_IF_END33]]
+// CHECK1:       omp_if.end33:
 // CHECK1-NEXT:    store i32 0, i32* [[NN]], align 4
-// CHECK1-NEXT:    [[TMP148:%.*]] = load i32, i32* [[NN]], align 4
-// CHECK1-NEXT:    [[CONV32:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP148]], i32* [[CONV32]], align 4
-// CHECK1-NEXT:    [[TMP149:%.*]] = load i64, i64* [[NN_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP149]], i64* [[TMP151]], align 8
-// CHECK1-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP149]], i64* [[TMP153]], align 8
-// CHECK1-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP154]], align 8
-// CHECK1-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP157:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154.region_id, i32 1, i8** [[TMP155]], i8** [[TMP156]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP158:%.*]] = icmp ne i32 [[TMP157]], 0
-// CHECK1-NEXT:    br i1 [[TMP158]], label [[OMP_OFFLOAD_FAILED36:%.*]], label [[OMP_OFFLOAD_CONT37:%.*]]
-// CHECK1:       omp_offload.failed36:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154(i64 [[TMP149]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT37]]
-// CHECK1:       omp_offload.cont37:
-// CHECK1-NEXT:    [[TMP159:%.*]] = load i32, i32* [[NN]], align 4
-// CHECK1-NEXT:    [[CONV39:%.*]] = bitcast i64* [[NN_CASTED38]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP159]], i32* [[CONV39]], align 4
-// CHECK1-NEXT:    [[TMP160:%.*]] = load i64, i64* [[NN_CASTED38]], align 8
-// CHECK1-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS40]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP162:%.*]] = bitcast i8** [[TMP161]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP160]], i64* [[TMP162]], align 8
-// CHECK1-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS41]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP164:%.*]] = bitcast i8** [[TMP163]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP160]], i64* [[TMP164]], align 8
-// CHECK1-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS42]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP165]], align 8
-// CHECK1-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS40]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS41]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP168:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157.region_id, i32 1, i8** [[TMP166]], i8** [[TMP167]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.18, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.19, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP169:%.*]] = icmp ne i32 [[TMP168]], 0
-// CHECK1-NEXT:    br i1 [[TMP169]], label [[OMP_OFFLOAD_FAILED43:%.*]], label [[OMP_OFFLOAD_CONT44:%.*]]
-// CHECK1:       omp_offload.failed43:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157(i64 [[TMP160]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT44]]
-// CHECK1:       omp_offload.cont44:
-// CHECK1-NEXT:    [[TMP170:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    [[TMP171:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP171]])
-// CHECK1-NEXT:    ret i32 [[TMP170]]
+// CHECK1-NEXT:    [[TMP172:%.*]] = load i32, i32* [[NN]], align 4
+// CHECK1-NEXT:    [[CONV34:%.*]] = bitcast i64* [[NN_CASTED]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP172]], i32* [[CONV34]], align 4
+// CHECK1-NEXT:    [[TMP173:%.*]] = load i64, i64* [[NN_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP174:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP175:%.*]] = bitcast i8** [[TMP174]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP173]], i64* [[TMP175]], align 8
+// CHECK1-NEXT:    [[TMP176:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP177:%.*]] = bitcast i8** [[TMP176]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP173]], i64* [[TMP177]], align 8
+// CHECK1-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS37]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP178]], align 8
+// CHECK1-NEXT:    [[TMP179:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS38:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP181:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP181]], align 4
+// CHECK1-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP182]], align 4
+// CHECK1-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP179]], i8*** [[TMP183]], align 8
+// CHECK1-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP180]], i8*** [[TMP184]], align 8
+// CHECK1-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP185]], align 8
+// CHECK1-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP186]], align 8
+// CHECK1-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP187]], align 8
+// CHECK1-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP188]], align 8
+// CHECK1-NEXT:    [[TMP189:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]])
+// CHECK1-NEXT:    [[TMP190:%.*]] = icmp ne i32 [[TMP189]], 0
+// CHECK1-NEXT:    br i1 [[TMP190]], label [[OMP_OFFLOAD_FAILED39:%.*]], label [[OMP_OFFLOAD_CONT40:%.*]]
+// CHECK1:       omp_offload.failed39:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154(i64 [[TMP173]]) #[[ATTR3]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT40]]
+// CHECK1:       omp_offload.cont40:
+// CHECK1-NEXT:    [[TMP191:%.*]] = load i32, i32* [[NN]], align 4
+// CHECK1-NEXT:    [[CONV42:%.*]] = bitcast i64* [[NN_CASTED41]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP191]], i32* [[CONV42]], align 4
+// CHECK1-NEXT:    [[TMP192:%.*]] = load i64, i64* [[NN_CASTED41]], align 8
+// CHECK1-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS43]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP194:%.*]] = bitcast i8** [[TMP193]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP192]], i64* [[TMP194]], align 8
+// CHECK1-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS44]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP196:%.*]] = bitcast i8** [[TMP195]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP192]], i64* [[TMP196]], align 8
+// CHECK1-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS45]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP197]], align 8
+// CHECK1-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS43]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS44]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS46:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP200]], align 4
+// CHECK1-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP201]], align 4
+// CHECK1-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP198]], i8*** [[TMP202]], align 8
+// CHECK1-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP199]], i8*** [[TMP203]], align 8
+// CHECK1-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.18, i32 0, i32 0), i64** [[TMP204]], align 8
+// CHECK1-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.19, i32 0, i32 0), i64** [[TMP205]], align 8
+// CHECK1-NEXT:    [[TMP206:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP206]], align 8
+// CHECK1-NEXT:    [[TMP207:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP207]], align 8
+// CHECK1-NEXT:    [[TMP208:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]])
+// CHECK1-NEXT:    [[TMP209:%.*]] = icmp ne i32 [[TMP208]], 0
+// CHECK1-NEXT:    br i1 [[TMP209]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
+// CHECK1:       omp_offload.failed47:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157(i64 [[TMP192]]) #[[ATTR3]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
+// CHECK1:       omp_offload.cont48:
+// CHECK1-NEXT:    [[TMP210:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP211:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP211]])
+// CHECK1-NEXT:    ret i32 [[TMP210]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
@@ -741,6 +826,7 @@ int bar(int n){
 // CHECK1-NEXT:    [[AA_CASTED_I:%.*]] = alloca i64, align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i64, align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED5_I:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK1-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -781,23 +867,39 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK1-NEXT:    [[TMP26:%.*]] = load i32, i32* [[TMP24]], align 4
-// CHECK1-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 [[TMP26]], i32 0, i8* null, i32 0, i8* null) #[[ATTR3]]
-// CHECK1-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK1-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP27]], align 4, !noalias !24
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP28]], align 4, !noalias !24
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP29]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP30]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP22]], i64** [[TMP31]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP32]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP33]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP34]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP25]], i32 [[TMP26]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR3]]
+// CHECK1-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK1-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK1:       omp_offload.failed.i:
-// CHECK1-NEXT:    [[TMP29:%.*]] = load i16, i16* [[TMP16]], align 2
+// CHECK1-NEXT:    [[TMP37:%.*]] = load i16, i16* [[TMP16]], align 2
 // CHECK1-NEXT:    [[CONV_I:%.*]] = bitcast i64* [[AA_CASTED_I]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP29]], i16* [[CONV_I]], align 2, !noalias !24
-// CHECK1-NEXT:    [[TMP30:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !24
-// CHECK1-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    store i16 [[TMP37]], i16* [[CONV_I]], align 2, !noalias !24
+// CHECK1-NEXT:    [[TMP38:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP39:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK1-NEXT:    [[CONV4_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED_I]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP31]], i32* [[CONV4_I]], align 4, !noalias !24
-// CHECK1-NEXT:    [[TMP32:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED_I]], align 8, !noalias !24
-// CHECK1-NEXT:    [[TMP33:%.*]] = load i32, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    store i32 [[TMP39]], i32* [[CONV4_I]], align 4, !noalias !24
+// CHECK1-NEXT:    [[TMP40:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED_I]], align 8, !noalias !24
+// CHECK1-NEXT:    [[TMP41:%.*]] = load i32, i32* [[TMP24]], align 4
 // CHECK1-NEXT:    [[CONV6_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED5_I]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP33]], i32* [[CONV6_I]], align 4, !noalias !24
-// CHECK1-NEXT:    [[TMP34:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5_I]], align 8, !noalias !24
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101(i64 [[TMP30]], i64 [[TMP32]], i64 [[TMP34]]) #[[ATTR3]]
+// CHECK1-NEXT:    store i32 [[TMP41]], i32* [[CONV6_I]], align 4, !noalias !24
+// CHECK1-NEXT:    [[TMP42:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5_I]], align 8, !noalias !24
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101(i64 [[TMP38]], i64 [[TMP40]], i64 [[TMP42]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
 // CHECK1:       .omp_outlined..1.exit:
 // CHECK1-NEXT:    ret i32 0
@@ -1137,9 +1239,26 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP6]], align 8
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.21, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.22, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.21, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.22, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182(i64 [[TMP1]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1287,9 +1406,26 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK1-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK1-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1299,15 +1435,15 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP42:%.*]] = mul nsw i64 1, [[TMP2]]
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP42]]
+// CHECK1-NEXT:    [[TMP50:%.*]] = mul nsw i64 1, [[TMP2]]
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP50]]
 // CHECK1-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
-// CHECK1-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK1-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP44]]
-// CHECK1-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK1-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP52]]
+// CHECK1-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK1-NEXT:    ret i32 [[ADD4]]
 //
 //
@@ -1379,9 +1515,26 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
 // CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.27, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.28, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK1-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.27, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.28, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK1-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1391,8 +1544,8 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP31]]
+// CHECK1-NEXT:    [[TMP39:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP39]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -1448,9 +1601,26 @@ int bar(int n){
 // CHECK1-NEXT:    store i8* null, i8** [[TMP19]], align 8
 // CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.30, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.31, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK1-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.30, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.31, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK1-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1460,8 +1630,8 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP24]]
+// CHECK1-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227
@@ -1698,20 +1868,20 @@ int bar(int n){
 // CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS12:%.*]] = alloca [2 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS13:%.*]] = alloca [2 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS14:%.*]] = alloca [2 x i8*], align 4
-// CHECK3-NEXT:    [[A_CASTED17:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [9 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [9 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [9 x i8*], align 4
+// CHECK3-NEXT:    [[A_CASTED18:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [9 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [9 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [9 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 4
 // CHECK3-NEXT:    [[NN:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[NN_CASTED:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[NN_CASTED32:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS29:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS30:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS31:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[NN_CASTED35:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS36:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS37:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS38:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK3-NEXT:    store i32 0, i32* [[A]], align 4
@@ -1816,204 +1986,289 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP63]], align 4
 // CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP64]], i8** [[TMP65]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
-// CHECK3-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK3-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK3-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP71]], align 4
+// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP73]], align 4
+// CHECK3-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK3-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i32 [[TMP58]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP68:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP68]], i32* [[A_CASTED9]], align 4
-// CHECK3-NEXT:    [[TMP69:%.*]] = load i32, i32* [[A_CASTED9]], align 4
-// CHECK3-NEXT:    [[TMP70:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK3-NEXT:    [[TMP76:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP76]], i32* [[A_CASTED9]], align 4
+// CHECK3-NEXT:    [[TMP77:%.*]] = load i32, i32* [[A_CASTED9]], align 4
+// CHECK3-NEXT:    [[TMP78:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK3-NEXT:    [[CONV11:%.*]] = bitcast i32* [[AA_CASTED10]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP70]], i16* [[CONV11]], align 2
-// CHECK3-NEXT:    [[TMP71:%.*]] = load i32, i32* [[AA_CASTED10]], align 4
-// CHECK3-NEXT:    [[TMP72:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP72]], 10
+// CHECK3-NEXT:    store i16 [[TMP78]], i16* [[CONV11]], align 2
+// CHECK3-NEXT:    [[TMP79:%.*]] = load i32, i32* [[AA_CASTED10]], align 4
+// CHECK3-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP80]], 10
 // CHECK3-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK3:       omp_if.then:
-// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP69]], i32* [[TMP74]], align 4
-// CHECK3-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP69]], i32* [[TMP76]], align 4
-// CHECK3-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP77]], align 4
-// CHECK3-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP71]], i32* [[TMP79]], align 4
-// CHECK3-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP71]], i32* [[TMP81]], align 4
-// CHECK3-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP82]], align 4
-// CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP85:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP83]], i8** [[TMP84]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP86:%.*]] = icmp ne i32 [[TMP85]], 0
-// CHECK3-NEXT:    br i1 [[TMP86]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK3:       omp_offload.failed15:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK3:       omp_offload.cont16:
+// CHECK3-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP77]], i32* [[TMP82]], align 4
+// CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP77]], i32* [[TMP84]], align 4
+// CHECK3-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP85]], align 4
+// CHECK3-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP79]], i32* [[TMP87]], align 4
+// CHECK3-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP79]], i32* [[TMP89]], align 4
+// CHECK3-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP90]], align 4
+// CHECK3-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP93]], align 4
+// CHECK3-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP94]], align 4
+// CHECK3-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP91]], i8*** [[TMP95]], align 4
+// CHECK3-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP92]], i8*** [[TMP96]], align 4
+// CHECK3-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP97]], align 4
+// CHECK3-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP98]], align 4
+// CHECK3-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP99]], align 4
+// CHECK3-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP100]], align 4
+// CHECK3-NEXT:    [[TMP101:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP102:%.*]] = icmp ne i32 [[TMP101]], 0
+// CHECK3-NEXT:    br i1 [[TMP102]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP77]], i32 [[TMP79]]) #[[ATTR3]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
 // CHECK3-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK3:       omp_if.else:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR3]]
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP77]], i32 [[TMP79]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP87:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP87]], i32* [[A_CASTED17]], align 4
-// CHECK3-NEXT:    [[TMP88:%.*]] = load i32, i32* [[A_CASTED17]], align 4
-// CHECK3-NEXT:    [[TMP89:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP18:%.*]] = icmp sgt i32 [[TMP89]], 20
-// CHECK3-NEXT:    br i1 [[CMP18]], label [[OMP_IF_THEN19:%.*]], label [[OMP_IF_ELSE25:%.*]]
-// CHECK3:       omp_if.then19:
-// CHECK3-NEXT:    [[TMP90:%.*]] = mul nuw i32 [[TMP1]], 4
-// CHECK3-NEXT:    [[TMP91:%.*]] = sext i32 [[TMP90]] to i64
-// CHECK3-NEXT:    [[TMP92:%.*]] = mul nuw i32 5, [[TMP3]]
-// CHECK3-NEXT:    [[TMP93:%.*]] = mul nuw i32 [[TMP92]], 8
-// CHECK3-NEXT:    [[TMP94:%.*]] = sext i32 [[TMP93]] to i64
-// CHECK3-NEXT:    [[TMP95:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP95]], i8* align 4 bitcast ([9 x i64]* @.offload_sizes.10 to i8*), i32 72, i1 false)
-// CHECK3-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP88]], i32* [[TMP97]], align 4
-// CHECK3-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP88]], i32* [[TMP99]], align 4
-// CHECK3-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP100]], align 4
-// CHECK3-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
-// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 4
-// CHECK3-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
-// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 4
-// CHECK3-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP105]], align 4
-// CHECK3-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP107]], align 4
-// CHECK3-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP109]], align 4
-// CHECK3-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP110]], align 4
-// CHECK3-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to float**
-// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP112]], align 4
-// CHECK3-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
-// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP114]], align 4
-// CHECK3-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK3-NEXT:    store i64 [[TMP91]], i64* [[TMP115]], align 4
-// CHECK3-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP103:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP103]], i32* [[A_CASTED18]], align 4
+// CHECK3-NEXT:    [[TMP104:%.*]] = load i32, i32* [[A_CASTED18]], align 4
+// CHECK3-NEXT:    [[TMP105:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP19:%.*]] = icmp sgt i32 [[TMP105]], 20
+// CHECK3-NEXT:    br i1 [[CMP19]], label [[OMP_IF_THEN20:%.*]], label [[OMP_IF_ELSE27:%.*]]
+// CHECK3:       omp_if.then20:
+// CHECK3-NEXT:    [[TMP106:%.*]] = mul nuw i32 [[TMP1]], 4
+// CHECK3-NEXT:    [[TMP107:%.*]] = sext i32 [[TMP106]] to i64
+// CHECK3-NEXT:    [[TMP108:%.*]] = mul nuw i32 5, [[TMP3]]
+// CHECK3-NEXT:    [[TMP109:%.*]] = mul nuw i32 [[TMP108]], 8
+// CHECK3-NEXT:    [[TMP110:%.*]] = sext i32 [[TMP109]] to i64
+// CHECK3-NEXT:    [[TMP111:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP111]], i8* align 4 bitcast ([9 x i64]* @.offload_sizes.10 to i8*), i32 72, i1 false)
+// CHECK3-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP104]], i32* [[TMP113]], align 4
+// CHECK3-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP104]], i32* [[TMP115]], align 4
+// CHECK3-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 0
 // CHECK3-NEXT:    store i8* null, i8** [[TMP116]], align 4
-// CHECK3-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to [5 x [10 x double]]**
-// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP118]], align 4
-// CHECK3-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
-// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 4
-// CHECK3-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to [10 x float]**
+// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP118]], align 4
+// CHECK3-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [10 x float]**
+// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP120]], align 4
+// CHECK3-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 1
 // CHECK3-NEXT:    store i8* null, i8** [[TMP121]], align 4
-// CHECK3-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP123:%.*]] = bitcast i8** [[TMP122]] to i32*
-// CHECK3-NEXT:    store i32 5, i32* [[TMP123]], align 4
-// CHECK3-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 5
+// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP123]], align 4
+// CHECK3-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP125:%.*]] = bitcast i8** [[TMP124]] to i32*
-// CHECK3-NEXT:    store i32 5, i32* [[TMP125]], align 4
-// CHECK3-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 5
+// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP125]], align 4
+// CHECK3-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 2
 // CHECK3-NEXT:    store i8* null, i8** [[TMP126]], align 4
-// CHECK3-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 6
-// CHECK3-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP128]], align 4
-// CHECK3-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 6
-// CHECK3-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP130]], align 4
-// CHECK3-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 6
-// CHECK3-NEXT:    store i8* null, i8** [[TMP131]], align 4
-// CHECK3-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 7
-// CHECK3-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to double**
-// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP133]], align 4
-// CHECK3-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 7
-// CHECK3-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to double**
-// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP135]], align 4
-// CHECK3-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK3-NEXT:    store i64 [[TMP94]], i64* [[TMP136]], align 4
-// CHECK3-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to float**
+// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP128]], align 4
+// CHECK3-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to float**
+// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP130]], align 4
+// CHECK3-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK3-NEXT:    store i64 [[TMP107]], i64* [[TMP131]], align 4
+// CHECK3-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 3
+// CHECK3-NEXT:    store i8* null, i8** [[TMP132]], align 4
+// CHECK3-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to [5 x [10 x double]]**
+// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP134]], align 4
+// CHECK3-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to [5 x [10 x double]]**
+// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP136]], align 4
+// CHECK3-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 4
 // CHECK3-NEXT:    store i8* null, i8** [[TMP137]], align 4
-// CHECK3-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 8
-// CHECK3-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to %struct.TT**
-// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP139]], align 4
-// CHECK3-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 8
-// CHECK3-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to %struct.TT**
-// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP141]], align 4
-// CHECK3-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32*
+// CHECK3-NEXT:    store i32 5, i32* [[TMP139]], align 4
+// CHECK3-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32*
+// CHECK3-NEXT:    store i32 5, i32* [[TMP141]], align 4
+// CHECK3-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 5
 // CHECK3-NEXT:    store i8* null, i8** [[TMP142]], align 4
-// CHECK3-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP146:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP143]], i8** [[TMP144]], i64* [[TMP145]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.11, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP147:%.*]] = icmp ne i32 [[TMP146]], 0
-// CHECK3-NEXT:    br i1 [[TMP147]], label [[OMP_OFFLOAD_FAILED23:%.*]], label [[OMP_OFFLOAD_CONT24:%.*]]
-// CHECK3:       omp_offload.failed23:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT24]]
-// CHECK3:       omp_offload.cont24:
-// CHECK3-NEXT:    br label [[OMP_IF_END26:%.*]]
-// CHECK3:       omp_if.else25:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_IF_END26]]
-// CHECK3:       omp_if.end26:
+// CHECK3-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 6
+// CHECK3-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP144]], align 4
+// CHECK3-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 6
+// CHECK3-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP146]], align 4
+// CHECK3-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 6
+// CHECK3-NEXT:    store i8* null, i8** [[TMP147]], align 4
+// CHECK3-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to double**
+// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP149]], align 4
+// CHECK3-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to double**
+// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP151]], align 4
+// CHECK3-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK3-NEXT:    store i64 [[TMP110]], i64* [[TMP152]], align 4
+// CHECK3-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 7
+// CHECK3-NEXT:    store i8* null, i8** [[TMP153]], align 4
+// CHECK3-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP155:%.*]] = bitcast i8** [[TMP154]] to %struct.TT**
+// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP155]], align 4
+// CHECK3-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to %struct.TT**
+// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP157]], align 4
+// CHECK3-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 8
+// CHECK3-NEXT:    store i8* null, i8** [[TMP158]], align 4
+// CHECK3-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS24:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP162]], align 4
+// CHECK3-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 9, i32* [[TMP163]], align 4
+// CHECK3-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP159]], i8*** [[TMP164]], align 4
+// CHECK3-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP160]], i8*** [[TMP165]], align 4
+// CHECK3-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP161]], i64** [[TMP166]], align 4
+// CHECK3-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.11, i32 0, i32 0), i64** [[TMP167]], align 4
+// CHECK3-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP168]], align 4
+// CHECK3-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP169]], align 4
+// CHECK3-NEXT:    [[TMP170:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS24]])
+// CHECK3-NEXT:    [[TMP171:%.*]] = icmp ne i32 [[TMP170]], 0
+// CHECK3-NEXT:    br i1 [[TMP171]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
+// CHECK3:       omp_offload.failed25:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP104]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
+// CHECK3:       omp_offload.cont26:
+// CHECK3-NEXT:    br label [[OMP_IF_END28:%.*]]
+// CHECK3:       omp_if.else27:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP104]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR3]]
+// CHECK3-NEXT:    br label [[OMP_IF_END28]]
+// CHECK3:       omp_if.end28:
 // CHECK3-NEXT:    store i32 0, i32* [[NN]], align 4
-// CHECK3-NEXT:    [[TMP148:%.*]] = load i32, i32* [[NN]], align 4
-// CHECK3-NEXT:    store i32 [[TMP148]], i32* [[NN_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP149:%.*]] = load i32, i32* [[NN_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP149]], i32* [[TMP151]], align 4
-// CHECK3-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP149]], i32* [[TMP153]], align 4
-// CHECK3-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP154]], align 4
-// CHECK3-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP157:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154.region_id, i32 1, i8** [[TMP155]], i8** [[TMP156]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP158:%.*]] = icmp ne i32 [[TMP157]], 0
-// CHECK3-NEXT:    br i1 [[TMP158]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK3:       omp_offload.failed30:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154(i32 [[TMP149]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK3:       omp_offload.cont31:
-// CHECK3-NEXT:    [[TMP159:%.*]] = load i32, i32* [[NN]], align 4
-// CHECK3-NEXT:    store i32 [[TMP159]], i32* [[NN_CASTED32]], align 4
-// CHECK3-NEXT:    [[TMP160:%.*]] = load i32, i32* [[NN_CASTED32]], align 4
-// CHECK3-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP162:%.*]] = bitcast i8** [[TMP161]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP160]], i32* [[TMP162]], align 4
-// CHECK3-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP164:%.*]] = bitcast i8** [[TMP163]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP160]], i32* [[TMP164]], align 4
-// CHECK3-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP165]], align 4
-// CHECK3-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP168:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157.region_id, i32 1, i8** [[TMP166]], i8** [[TMP167]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.18, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.19, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP169:%.*]] = icmp ne i32 [[TMP168]], 0
-// CHECK3-NEXT:    br i1 [[TMP169]], label [[OMP_OFFLOAD_FAILED36:%.*]], label [[OMP_OFFLOAD_CONT37:%.*]]
-// CHECK3:       omp_offload.failed36:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157(i32 [[TMP160]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT37]]
-// CHECK3:       omp_offload.cont37:
-// CHECK3-NEXT:    [[TMP170:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    [[TMP171:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP171]])
-// CHECK3-NEXT:    ret i32 [[TMP170]]
+// CHECK3-NEXT:    [[TMP172:%.*]] = load i32, i32* [[NN]], align 4
+// CHECK3-NEXT:    store i32 [[TMP172]], i32* [[NN_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP173:%.*]] = load i32, i32* [[NN_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP174:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP175:%.*]] = bitcast i8** [[TMP174]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP173]], i32* [[TMP175]], align 4
+// CHECK3-NEXT:    [[TMP176:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP177:%.*]] = bitcast i8** [[TMP176]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP173]], i32* [[TMP177]], align 4
+// CHECK3-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP178]], align 4
+// CHECK3-NEXT:    [[TMP179:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS32:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP181:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP181]], align 4
+// CHECK3-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP182]], align 4
+// CHECK3-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP179]], i8*** [[TMP183]], align 4
+// CHECK3-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP180]], i8*** [[TMP184]], align 4
+// CHECK3-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP185]], align 4
+// CHECK3-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP186]], align 4
+// CHECK3-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP187]], align 4
+// CHECK3-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP188]], align 4
+// CHECK3-NEXT:    [[TMP189:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]])
+// CHECK3-NEXT:    [[TMP190:%.*]] = icmp ne i32 [[TMP189]], 0
+// CHECK3-NEXT:    br i1 [[TMP190]], label [[OMP_OFFLOAD_FAILED33:%.*]], label [[OMP_OFFLOAD_CONT34:%.*]]
+// CHECK3:       omp_offload.failed33:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l154(i32 [[TMP173]]) #[[ATTR3]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT34]]
+// CHECK3:       omp_offload.cont34:
+// CHECK3-NEXT:    [[TMP191:%.*]] = load i32, i32* [[NN]], align 4
+// CHECK3-NEXT:    store i32 [[TMP191]], i32* [[NN_CASTED35]], align 4
+// CHECK3-NEXT:    [[TMP192:%.*]] = load i32, i32* [[NN_CASTED35]], align 4
+// CHECK3-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP194:%.*]] = bitcast i8** [[TMP193]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP192]], i32* [[TMP194]], align 4
+// CHECK3-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP196:%.*]] = bitcast i8** [[TMP195]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP192]], i32* [[TMP196]], align 4
+// CHECK3-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP197]], align 4
+// CHECK3-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS39:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS39]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP200]], align 4
+// CHECK3-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS39]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP201]], align 4
+// CHECK3-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS39]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP198]], i8*** [[TMP202]], align 4
+// CHECK3-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS39]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP199]], i8*** [[TMP203]], align 4
+// CHECK3-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS39]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.18, i32 0, i32 0), i64** [[TMP204]], align 4
+// CHECK3-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS39]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.19, i32 0, i32 0), i64** [[TMP205]], align 4
+// CHECK3-NEXT:    [[TMP206:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS39]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP206]], align 4
+// CHECK3-NEXT:    [[TMP207:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS39]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP207]], align 4
+// CHECK3-NEXT:    [[TMP208:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS39]])
+// CHECK3-NEXT:    [[TMP209:%.*]] = icmp ne i32 [[TMP208]], 0
+// CHECK3-NEXT:    br i1 [[TMP209]], label [[OMP_OFFLOAD_FAILED40:%.*]], label [[OMP_OFFLOAD_CONT41:%.*]]
+// CHECK3:       omp_offload.failed40:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l157(i32 [[TMP192]]) #[[ATTR3]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT41]]
+// CHECK3:       omp_offload.cont41:
+// CHECK3-NEXT:    [[TMP210:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    [[TMP211:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP211]])
+// CHECK3-NEXT:    ret i32 [[TMP210]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101
@@ -2097,6 +2352,7 @@ int bar(int n){
 // CHECK3-NEXT:    [[AA_CASTED_I:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED4_I:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK3-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 4
 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -2137,21 +2393,37 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK3-NEXT:    [[TMP26:%.*]] = load i32, i32* [[TMP24]], align 4
-// CHECK3-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 [[TMP26]], i32 0, i8* null, i32 0, i8* null) #[[ATTR3]]
-// CHECK3-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK3-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP27]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP28]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP20]], i8*** [[TMP29]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP30]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP22]], i64** [[TMP31]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP32]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP33]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP34]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP25]], i32 [[TMP26]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR3]]
+// CHECK3-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK3-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK3:       omp_offload.failed.i:
-// CHECK3-NEXT:    [[TMP29:%.*]] = load i16, i16* [[TMP16]], align 2
+// CHECK3-NEXT:    [[TMP37:%.*]] = load i16, i16* [[TMP16]], align 2
 // CHECK3-NEXT:    [[CONV_I:%.*]] = bitcast i32* [[AA_CASTED_I]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP29]], i16* [[CONV_I]], align 2, !noalias !25
-// CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !25
-// CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP23]], align 4
-// CHECK3-NEXT:    store i32 [[TMP31]], i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !25
-// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !25
-// CHECK3-NEXT:    [[TMP33:%.*]] = load i32, i32* [[TMP24]], align 4
-// CHECK3-NEXT:    store i32 [[TMP33]], i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !25
-// CHECK3-NEXT:    [[TMP34:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !25
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101(i32 [[TMP30]], i32 [[TMP32]], i32 [[TMP34]]) #[[ATTR3]]
+// CHECK3-NEXT:    store i16 [[TMP37]], i16* [[CONV_I]], align 2, !noalias !25
+// CHECK3-NEXT:    [[TMP38:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP39:%.*]] = load i32, i32* [[TMP23]], align 4
+// CHECK3-NEXT:    store i32 [[TMP39]], i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP40:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP41:%.*]] = load i32, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    store i32 [[TMP41]], i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !25
+// CHECK3-NEXT:    [[TMP42:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !25
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l101(i32 [[TMP38]], i32 [[TMP40]], i32 [[TMP42]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
 // CHECK3:       .omp_outlined..1.exit:
 // CHECK3-NEXT:    ret i32 0
@@ -2473,9 +2745,26 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP5]], align 4
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.21, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.22, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
-// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.21, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.22, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
+// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z6bazzzziPi_l182(i32 [[TMP0]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2622,9 +2911,26 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK3-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK3-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2634,15 +2940,15 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP42:%.*]] = mul nsw i32 1, [[TMP1]]
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP42]]
+// CHECK3-NEXT:    [[TMP50:%.*]] = mul nsw i32 1, [[TMP1]]
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP50]]
 // CHECK3-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
-// CHECK3-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK3-NEXT:    [[CONV:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK3-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP44]]
-// CHECK3-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK3-NEXT:    [[CONV:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK3-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP52]]
+// CHECK3-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK3-NEXT:    ret i32 [[ADD3]]
 //
 //
@@ -2713,9 +3019,26 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP26]], align 4
 // CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.27, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.28, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK3-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.27, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.28, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK3-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2725,8 +3048,8 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l209(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP31]]
+// CHECK3-NEXT:    [[TMP39:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP39]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -2781,9 +3104,26 @@ int bar(int n){
 // CHECK3-NEXT:    store i8* null, i8** [[TMP19]], align 4
 // CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.30, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.31, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK3-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.30, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.31, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK3-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2793,8 +3133,8 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l192(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP24]]
+// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l227

diff --git a/clang/test/OpenMP/target_teams_depend_codegen.cpp b/clang/test/OpenMP/target_teams_depend_codegen.cpp
index 8117498..7011d27 100644 (file)
--- a/clang/test/OpenMP/target_teams_depend_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_depend_codegen.cpp
@@ -186,7 +186,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 0
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 0, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // CHECK:       [[FAIL]]
@@ -204,7 +204,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 2
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 2, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SZT]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT]], i32 0, i32 0), i8** null, i8** null, i32 0, i32 0, i32 0, i8* null, i32 0, i8* null)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 0, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 
 // CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]

diff --git a/clang/test/OpenMP/target_teams_distribute_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_codegen.cpp
index c99a8f0..5a6d1f9 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_codegen.cpp
@@ -327,14 +327,14 @@ int bar(int n){
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS17:%.*]] = alloca [2 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS18:%.*]] = alloca [2 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP19:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[A_CASTED23:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED25:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS29:%.*]] = alloca [10 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS30:%.*]] = alloca [10 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS31:%.*]] = alloca [10 x i8*], align 8
+// CHECK1-NEXT:    [[DOTCAPTURE_EXPR_23:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[A_CASTED24:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED26:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS30:%.*]] = alloca [10 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS31:%.*]] = alloca [10 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS32:%.*]] = alloca [10 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [10 x i64], align 8
-// CHECK1-NEXT:    [[_TMP32:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[_TMP33:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK1-NEXT:    store i32 0, i32* [[A]], align 4
@@ -445,179 +445,230 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK1-NEXT:    [[TMP68:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l113.region_id, i32 1, i8** [[TMP66]], i8** [[TMP67]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP69:%.*]] = icmp ne i32 [[TMP68]], 0
-// CHECK1-NEXT:    br i1 [[TMP69]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK1-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP69]], align 4
+// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP66]], i8*** [[TMP70]], align 8
+// CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP67]], i8*** [[TMP71]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 8
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP75]], align 8
+// CHECK1-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l113.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK1-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l113(i64 [[TMP60]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP70:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP78:%.*]] = load i32, i32* [[A]], align 4
 // CHECK1-NEXT:    [[CONV13:%.*]] = bitcast i64* [[A_CASTED12]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP70]], i32* [[CONV13]], align 4
-// CHECK1-NEXT:    [[TMP71:%.*]] = load i64, i64* [[A_CASTED12]], align 8
-// CHECK1-NEXT:    [[TMP72:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK1-NEXT:    store i32 [[TMP78]], i32* [[CONV13]], align 4
+// CHECK1-NEXT:    [[TMP79:%.*]] = load i64, i64* [[A_CASTED12]], align 8
+// CHECK1-NEXT:    [[TMP80:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK1-NEXT:    [[CONV15:%.*]] = bitcast i64* [[AA_CASTED14]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP72]], i16* [[CONV15]], align 2
-// CHECK1-NEXT:    [[TMP73:%.*]] = load i64, i64* [[AA_CASTED14]], align 8
-// CHECK1-NEXT:    [[TMP74:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP74]], 10
+// CHECK1-NEXT:    store i16 [[TMP80]], i16* [[CONV15]], align 2
+// CHECK1-NEXT:    [[TMP81:%.*]] = load i64, i64* [[AA_CASTED14]], align 8
+// CHECK1-NEXT:    [[TMP82:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP82]], 10
 // CHECK1-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK1:       omp_if.then:
-// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP71]], i64* [[TMP76]], align 8
-// CHECK1-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP71]], i64* [[TMP78]], align 8
-// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP79]], align 8
-// CHECK1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP73]], i64* [[TMP81]], align 8
-// CHECK1-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP73]], i64* [[TMP83]], align 8
-// CHECK1-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP84]], align 8
-// CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP79]], i64* [[TMP84]], align 8
+// CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP79]], i64* [[TMP86]], align 8
+// CHECK1-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP87]], align 8
+// CHECK1-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP81]], i64* [[TMP89]], align 8
+// CHECK1-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP81]], i64* [[TMP91]], align 8
+// CHECK1-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP92]], align 8
+// CHECK1-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK1-NEXT:    [[TMP87:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l120.region_id, i32 2, i8** [[TMP85]], i8** [[TMP86]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP88:%.*]] = icmp ne i32 [[TMP87]], 0
-// CHECK1-NEXT:    br i1 [[TMP88]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
-// CHECK1:       omp_offload.failed20:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l120(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
-// CHECK1:       omp_offload.cont21:
+// CHECK1-NEXT:    [[KERNEL_ARGS20:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP95]], align 4
+// CHECK1-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP96]], align 4
+// CHECK1-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP93]], i8*** [[TMP97]], align 8
+// CHECK1-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP94]], i8*** [[TMP98]], align 8
+// CHECK1-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP99]], align 8
+// CHECK1-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP100]], align 8
+// CHECK1-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP101]], align 8
+// CHECK1-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP102]], align 8
+// CHECK1-NEXT:    [[TMP103:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l120.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]])
+// CHECK1-NEXT:    [[TMP104:%.*]] = icmp ne i32 [[TMP103]], 0
+// CHECK1-NEXT:    br i1 [[TMP104]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
+// CHECK1:       omp_offload.failed21:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l120(i64 [[TMP79]], i64 [[TMP81]]) #[[ATTR3]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
+// CHECK1:       omp_offload.cont22:
 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK1:       omp_if.else:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l120(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR3]]
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l120(i64 [[TMP79]], i64 [[TMP81]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP89:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    store i32 [[TMP89]], i32* [[DOTCAPTURE_EXPR_22]], align 4
-// CHECK1-NEXT:    [[TMP90:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    [[CONV24:%.*]] = bitcast i64* [[A_CASTED23]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP90]], i32* [[CONV24]], align 4
-// CHECK1-NEXT:    [[TMP91:%.*]] = load i64, i64* [[A_CASTED23]], align 8
-// CHECK1-NEXT:    [[TMP92:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
-// CHECK1-NEXT:    [[CONV26:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED25]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP92]], i32* [[CONV26]], align 4
-// CHECK1-NEXT:    [[TMP93:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED25]], align 8
-// CHECK1-NEXT:    [[TMP94:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP27:%.*]] = icmp sgt i32 [[TMP94]], 20
-// CHECK1-NEXT:    br i1 [[CMP27]], label [[OMP_IF_THEN28:%.*]], label [[OMP_IF_ELSE35:%.*]]
-// CHECK1:       omp_if.then28:
-// CHECK1-NEXT:    [[TMP95:%.*]] = mul nuw i64 [[TMP2]], 4
-// CHECK1-NEXT:    [[TMP96:%.*]] = mul nuw i64 5, [[TMP5]]
-// CHECK1-NEXT:    [[TMP97:%.*]] = mul nuw i64 [[TMP96]], 8
-// CHECK1-NEXT:    [[TMP98:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP98]], i8* align 8 bitcast ([10 x i64]* @.offload_sizes.10 to i8*), i64 80, i1 false)
-// CHECK1-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP100:%.*]] = bitcast i8** [[TMP99]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP91]], i64* [[TMP100]], align 8
-// CHECK1-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP91]], i64* [[TMP102]], align 8
-// CHECK1-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP103]], align 8
-// CHECK1-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to [10 x float]**
-// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP105]], align 8
-// CHECK1-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to [10 x float]**
-// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP107]], align 8
-// CHECK1-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP108]], align 8
-// CHECK1-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP110]], align 8
-// CHECK1-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP112]], align 8
-// CHECK1-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP113]], align 8
-// CHECK1-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to float**
-// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP115]], align 8
-// CHECK1-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to float**
-// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP117]], align 8
-// CHECK1-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK1-NEXT:    store i64 [[TMP95]], i64* [[TMP118]], align 8
-// CHECK1-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i64 0, i64 3
+// CHECK1-NEXT:    [[TMP105:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    store i32 [[TMP105]], i32* [[DOTCAPTURE_EXPR_23]], align 4
+// CHECK1-NEXT:    [[TMP106:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[CONV25:%.*]] = bitcast i64* [[A_CASTED24]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP106]], i32* [[CONV25]], align 4
+// CHECK1-NEXT:    [[TMP107:%.*]] = load i64, i64* [[A_CASTED24]], align 8
+// CHECK1-NEXT:    [[TMP108:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_23]], align 4
+// CHECK1-NEXT:    [[CONV27:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED26]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP108]], i32* [[CONV27]], align 4
+// CHECK1-NEXT:    [[TMP109:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED26]], align 8
+// CHECK1-NEXT:    [[TMP110:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP28:%.*]] = icmp sgt i32 [[TMP110]], 20
+// CHECK1-NEXT:    br i1 [[CMP28]], label [[OMP_IF_THEN29:%.*]], label [[OMP_IF_ELSE37:%.*]]
+// CHECK1:       omp_if.then29:
+// CHECK1-NEXT:    [[TMP111:%.*]] = mul nuw i64 [[TMP2]], 4
+// CHECK1-NEXT:    [[TMP112:%.*]] = mul nuw i64 5, [[TMP5]]
+// CHECK1-NEXT:    [[TMP113:%.*]] = mul nuw i64 [[TMP112]], 8
+// CHECK1-NEXT:    [[TMP114:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP114]], i8* align 8 bitcast ([10 x i64]* @.offload_sizes.10 to i8*), i64 80, i1 false)
+// CHECK1-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP107]], i64* [[TMP116]], align 8
+// CHECK1-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP107]], i64* [[TMP118]], align 8
+// CHECK1-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 0
 // CHECK1-NEXT:    store i8* null, i8** [[TMP119]], align 8
-// CHECK1-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP121:%.*]] = bitcast i8** [[TMP120]] to [5 x [10 x double]]**
-// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP121]], align 8
-// CHECK1-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP123:%.*]] = bitcast i8** [[TMP122]] to [5 x [10 x double]]**
-// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP123]], align 8
-// CHECK1-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i64 0, i64 4
+// CHECK1-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP121:%.*]] = bitcast i8** [[TMP120]] to [10 x float]**
+// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP121]], align 8
+// CHECK1-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP123:%.*]] = bitcast i8** [[TMP122]] to [10 x float]**
+// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP123]], align 8
+// CHECK1-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 1
 // CHECK1-NEXT:    store i8* null, i8** [[TMP124]], align 8
-// CHECK1-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i64*
-// CHECK1-NEXT:    store i64 5, i64* [[TMP126]], align 8
-// CHECK1-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 5
+// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP126]], align 8
+// CHECK1-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i64*
-// CHECK1-NEXT:    store i64 5, i64* [[TMP128]], align 8
-// CHECK1-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i64 0, i64 5
+// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP128]], align 8
+// CHECK1-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 2
 // CHECK1-NEXT:    store i8* null, i8** [[TMP129]], align 8
-// CHECK1-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP131]], align 8
-// CHECK1-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP133]], align 8
-// CHECK1-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i64 0, i64 6
-// CHECK1-NEXT:    store i8* null, i8** [[TMP134]], align 8
-// CHECK1-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to double**
-// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP136]], align 8
-// CHECK1-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to double**
-// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP138]], align 8
-// CHECK1-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK1-NEXT:    store i64 [[TMP97]], i64* [[TMP139]], align 8
-// CHECK1-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i64 0, i64 7
+// CHECK1-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to float**
+// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP131]], align 8
+// CHECK1-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to float**
+// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP133]], align 8
+// CHECK1-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK1-NEXT:    store i64 [[TMP111]], i64* [[TMP134]], align 8
+// CHECK1-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 3
+// CHECK1-NEXT:    store i8* null, i8** [[TMP135]], align 8
+// CHECK1-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to [5 x [10 x double]]**
+// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP137]], align 8
+// CHECK1-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to [5 x [10 x double]]**
+// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP139]], align 8
+// CHECK1-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 4
 // CHECK1-NEXT:    store i8* null, i8** [[TMP140]], align 8
-// CHECK1-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to %struct.TT**
-// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP142]], align 8
-// CHECK1-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to %struct.TT**
-// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP144]], align 8
-// CHECK1-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i64 0, i64 8
+// CHECK1-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i64*
+// CHECK1-NEXT:    store i64 5, i64* [[TMP142]], align 8
+// CHECK1-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i64*
+// CHECK1-NEXT:    store i64 5, i64* [[TMP144]], align 8
+// CHECK1-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 5
 // CHECK1-NEXT:    store i8* null, i8** [[TMP145]], align 8
-// CHECK1-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 9
+// CHECK1-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 6
 // CHECK1-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP93]], i64* [[TMP147]], align 8
-// CHECK1-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 9
+// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP147]], align 8
+// CHECK1-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 6
 // CHECK1-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP93]], i64* [[TMP149]], align 8
-// CHECK1-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i64 0, i64 9
+// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP149]], align 8
+// CHECK1-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 6
 // CHECK1-NEXT:    store i8* null, i8** [[TMP150]], align 8
-// CHECK1-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to double**
+// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP152]], align 8
+// CHECK1-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to double**
+// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP154]], align 8
+// CHECK1-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK1-NEXT:    store i64 [[TMP113]], i64* [[TMP155]], align 8
+// CHECK1-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 7
+// CHECK1-NEXT:    store i8* null, i8** [[TMP156]], align 8
+// CHECK1-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP158:%.*]] = bitcast i8** [[TMP157]] to %struct.TT**
+// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP158]], align 8
+// CHECK1-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP160:%.*]] = bitcast i8** [[TMP159]] to %struct.TT**
+// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP160]], align 8
+// CHECK1-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 8
+// CHECK1-NEXT:    store i8* null, i8** [[TMP161]], align 8
+// CHECK1-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 9
+// CHECK1-NEXT:    [[TMP163:%.*]] = bitcast i8** [[TMP162]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP109]], i64* [[TMP163]], align 8
+// CHECK1-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 9
+// CHECK1-NEXT:    [[TMP165:%.*]] = bitcast i8** [[TMP164]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP109]], i64* [[TMP165]], align 8
+// CHECK1-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 9
+// CHECK1-NEXT:    store i8* null, i8** [[TMP166]], align 8
+// CHECK1-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK1-NEXT:    [[TMP154:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l145.region_id, i32 10, i8** [[TMP151]], i8** [[TMP152]], i64* [[TMP153]], i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.11, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP155:%.*]] = icmp ne i32 [[TMP154]], 0
-// CHECK1-NEXT:    br i1 [[TMP155]], label [[OMP_OFFLOAD_FAILED33:%.*]], label [[OMP_OFFLOAD_CONT34:%.*]]
-// CHECK1:       omp_offload.failed33:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l145(i64 [[TMP91]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP93]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT34]]
-// CHECK1:       omp_offload.cont34:
-// CHECK1-NEXT:    br label [[OMP_IF_END36:%.*]]
-// CHECK1:       omp_if.else35:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l145(i64 [[TMP91]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP93]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_IF_END36]]
-// CHECK1:       omp_if.end36:
-// CHECK1-NEXT:    [[TMP156:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    [[TMP157:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP157]])
-// CHECK1-NEXT:    ret i32 [[TMP156]]
+// CHECK1-NEXT:    [[KERNEL_ARGS34:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP170]], align 4
+// CHECK1-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 10, i32* [[TMP171]], align 4
+// CHECK1-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP167]], i8*** [[TMP172]], align 8
+// CHECK1-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP168]], i8*** [[TMP173]], align 8
+// CHECK1-NEXT:    [[TMP174:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP169]], i64** [[TMP174]], align 8
+// CHECK1-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.11, i32 0, i32 0), i64** [[TMP175]], align 8
+// CHECK1-NEXT:    [[TMP176:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP176]], align 8
+// CHECK1-NEXT:    [[TMP177:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP177]], align 8
+// CHECK1-NEXT:    [[TMP178:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l145.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]])
+// CHECK1-NEXT:    [[TMP179:%.*]] = icmp ne i32 [[TMP178]], 0
+// CHECK1-NEXT:    br i1 [[TMP179]], label [[OMP_OFFLOAD_FAILED35:%.*]], label [[OMP_OFFLOAD_CONT36:%.*]]
+// CHECK1:       omp_offload.failed35:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l145(i64 [[TMP107]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP109]]) #[[ATTR3]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT36]]
+// CHECK1:       omp_offload.cont36:
+// CHECK1-NEXT:    br label [[OMP_IF_END38:%.*]]
+// CHECK1:       omp_if.else37:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l145(i64 [[TMP107]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]], i64 [[TMP109]]) #[[ATTR3]]
+// CHECK1-NEXT:    br label [[OMP_IF_END38]]
+// CHECK1:       omp_if.end38:
+// CHECK1-NEXT:    [[TMP180:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP181:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP181]])
+// CHECK1-NEXT:    ret i32 [[TMP180]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103
@@ -753,6 +804,7 @@ int bar(int n){
 // CHECK1-NEXT:    [[AA_CASTED_I:%.*]] = alloca i64, align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i64, align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED5_I:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK1-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -794,23 +846,39 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK1-NEXT:    [[TMP26:%.*]] = load i32, i32* [[TMP24]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10) #[[ATTR3]]
-// CHECK1-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 [[TMP26]], i32 0, i8* null, i32 0, i8* null) #[[ATTR3]]
-// CHECK1-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK1-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP27]], align 4, !noalias !21
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP28]], align 4, !noalias !21
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP29]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP30]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP22]], i64** [[TMP31]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP32]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP33]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP34]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP25]], i32 [[TMP26]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR3]]
+// CHECK1-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK1-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK1:       omp_offload.failed.i:
-// CHECK1-NEXT:    [[TMP29:%.*]] = load i16, i16* [[TMP16]], align 2
+// CHECK1-NEXT:    [[TMP37:%.*]] = load i16, i16* [[TMP16]], align 2
 // CHECK1-NEXT:    [[CONV_I:%.*]] = bitcast i64* [[AA_CASTED_I]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP29]], i16* [[CONV_I]], align 2, !noalias !21
-// CHECK1-NEXT:    [[TMP30:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !21
-// CHECK1-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    store i16 [[TMP37]], i16* [[CONV_I]], align 2, !noalias !21
+// CHECK1-NEXT:    [[TMP38:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP39:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK1-NEXT:    [[CONV4_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED_I]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP31]], i32* [[CONV4_I]], align 4, !noalias !21
-// CHECK1-NEXT:    [[TMP32:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED_I]], align 8, !noalias !21
-// CHECK1-NEXT:    [[TMP33:%.*]] = load i32, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    store i32 [[TMP39]], i32* [[CONV4_I]], align 4, !noalias !21
+// CHECK1-NEXT:    [[TMP40:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED_I]], align 8, !noalias !21
+// CHECK1-NEXT:    [[TMP41:%.*]] = load i32, i32* [[TMP24]], align 4
 // CHECK1-NEXT:    [[CONV6_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED5_I]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP33]], i32* [[CONV6_I]], align 4, !noalias !21
-// CHECK1-NEXT:    [[TMP34:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5_I]], align 8, !noalias !21
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103(i64 [[TMP30]], i64 [[TMP32]], i64 [[TMP34]]) #[[ATTR3]]
+// CHECK1-NEXT:    store i32 [[TMP41]], i32* [[CONV6_I]], align 4, !noalias !21
+// CHECK1-NEXT:    [[TMP42:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5_I]], align 8, !noalias !21
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103(i64 [[TMP38]], i64 [[TMP40]], i64 [[TMP42]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
 // CHECK1:       .omp_outlined..1.exit:
 // CHECK1-NEXT:    ret i32 0
@@ -1386,9 +1454,26 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK1-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l218.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK1-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.14, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l218.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK1-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l218(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1398,15 +1483,15 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l218(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP42:%.*]] = mul nsw i64 1, [[TMP2]]
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP42]]
+// CHECK1-NEXT:    [[TMP50:%.*]] = mul nsw i64 1, [[TMP2]]
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP50]]
 // CHECK1-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
-// CHECK1-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK1-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP44]]
-// CHECK1-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK1-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP52]]
+// CHECK1-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK1-NEXT:    ret i32 [[ADD4]]
 //
 //
@@ -1511,9 +1596,26 @@ int bar(int n){
 // CHECK1-NEXT:    [[ADD8:%.*]] = add i32 [[TMP40]], 1
 // CHECK1-NEXT:    [[TMP41:%.*]] = zext i32 [[ADD8]] to i64
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP41]])
-// CHECK1-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l200.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK1-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP43]], align 4
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP34]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP35]], i8*** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP49]], align 8
+// CHECK1-NEXT:    [[TMP50:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l200.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP51:%.*]] = icmp ne i32 [[TMP50]], 0
+// CHECK1-NEXT:    br i1 [[TMP51]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l200(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], i64 [[TMP7]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1523,8 +1625,8 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l200(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], i64 [[TMP7]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP44:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP44]]
+// CHECK1-NEXT:    [[TMP52:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP52]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -1582,9 +1684,26 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK1-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l183.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.19, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.20, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK1-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.19, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.20, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l183.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK1-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l183(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1594,8 +1713,8 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l183(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP24]]
+// CHECK1-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l218
@@ -2027,14 +2146,14 @@ int bar(int n){
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS13:%.*]] = alloca [2 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS14:%.*]] = alloca [2 x i8*], align 4
 // CHECK3-NEXT:    [[_TMP15:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[A_CASTED19:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED20:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [10 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [10 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [10 x i8*], align 4
+// CHECK3-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[A_CASTED20:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED21:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [10 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [10 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [10 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [10 x i64], align 4
-// CHECK3-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK3-NEXT:    store i32 0, i32* [[A]], align 4
@@ -2140,178 +2259,229 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK3-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l113.region_id, i32 1, i8** [[TMP64]], i8** [[TMP65]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
-// CHECK3-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK3-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK3-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP71]], align 4
+// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP73]], align 4
+// CHECK3-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l113.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK3-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l113(i32 [[TMP58]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP68:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP68]], i32* [[A_CASTED9]], align 4
-// CHECK3-NEXT:    [[TMP69:%.*]] = load i32, i32* [[A_CASTED9]], align 4
-// CHECK3-NEXT:    [[TMP70:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK3-NEXT:    [[TMP76:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP76]], i32* [[A_CASTED9]], align 4
+// CHECK3-NEXT:    [[TMP77:%.*]] = load i32, i32* [[A_CASTED9]], align 4
+// CHECK3-NEXT:    [[TMP78:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK3-NEXT:    [[CONV11:%.*]] = bitcast i32* [[AA_CASTED10]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP70]], i16* [[CONV11]], align 2
-// CHECK3-NEXT:    [[TMP71:%.*]] = load i32, i32* [[AA_CASTED10]], align 4
-// CHECK3-NEXT:    [[TMP72:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP72]], 10
+// CHECK3-NEXT:    store i16 [[TMP78]], i16* [[CONV11]], align 2
+// CHECK3-NEXT:    [[TMP79:%.*]] = load i32, i32* [[AA_CASTED10]], align 4
+// CHECK3-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP80]], 10
 // CHECK3-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK3:       omp_if.then:
-// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP69]], i32* [[TMP74]], align 4
-// CHECK3-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP69]], i32* [[TMP76]], align 4
-// CHECK3-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP77]], align 4
-// CHECK3-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP71]], i32* [[TMP79]], align 4
-// CHECK3-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP71]], i32* [[TMP81]], align 4
-// CHECK3-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP82]], align 4
-// CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP77]], i32* [[TMP82]], align 4
+// CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP77]], i32* [[TMP84]], align 4
+// CHECK3-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP85]], align 4
+// CHECK3-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP79]], i32* [[TMP87]], align 4
+// CHECK3-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP79]], i32* [[TMP89]], align 4
+// CHECK3-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP90]], align 4
+// CHECK3-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK3-NEXT:    [[TMP85:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l120.region_id, i32 2, i8** [[TMP83]], i8** [[TMP84]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP86:%.*]] = icmp ne i32 [[TMP85]], 0
-// CHECK3-NEXT:    br i1 [[TMP86]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK3:       omp_offload.failed16:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l120(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP93]], align 4
+// CHECK3-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP94]], align 4
+// CHECK3-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP91]], i8*** [[TMP95]], align 4
+// CHECK3-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP92]], i8*** [[TMP96]], align 4
+// CHECK3-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP97]], align 4
+// CHECK3-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP98]], align 4
+// CHECK3-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP99]], align 4
+// CHECK3-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP100]], align 4
+// CHECK3-NEXT:    [[TMP101:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l120.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK3-NEXT:    [[TMP102:%.*]] = icmp ne i32 [[TMP101]], 0
+// CHECK3-NEXT:    br i1 [[TMP102]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK3:       omp_offload.failed17:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l120(i32 [[TMP77]], i32 [[TMP79]]) #[[ATTR3]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK3:       omp_offload.cont18:
 // CHECK3-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK3:       omp_if.else:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l120(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR3]]
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l120(i32 [[TMP77]], i32 [[TMP79]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP87:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    store i32 [[TMP87]], i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK3-NEXT:    [[TMP88:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP88]], i32* [[A_CASTED19]], align 4
-// CHECK3-NEXT:    [[TMP89:%.*]] = load i32, i32* [[A_CASTED19]], align 4
-// CHECK3-NEXT:    [[TMP90:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK3-NEXT:    store i32 [[TMP90]], i32* [[DOTCAPTURE_EXPR__CASTED20]], align 4
-// CHECK3-NEXT:    [[TMP91:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED20]], align 4
-// CHECK3-NEXT:    [[TMP92:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP21:%.*]] = icmp sgt i32 [[TMP92]], 20
-// CHECK3-NEXT:    br i1 [[CMP21]], label [[OMP_IF_THEN22:%.*]], label [[OMP_IF_ELSE29:%.*]]
-// CHECK3:       omp_if.then22:
-// CHECK3-NEXT:    [[TMP93:%.*]] = mul nuw i32 [[TMP1]], 4
-// CHECK3-NEXT:    [[TMP94:%.*]] = sext i32 [[TMP93]] to i64
-// CHECK3-NEXT:    [[TMP95:%.*]] = mul nuw i32 5, [[TMP3]]
-// CHECK3-NEXT:    [[TMP96:%.*]] = mul nuw i32 [[TMP95]], 8
-// CHECK3-NEXT:    [[TMP97:%.*]] = sext i32 [[TMP96]] to i64
-// CHECK3-NEXT:    [[TMP98:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP98]], i8* align 4 bitcast ([10 x i64]* @.offload_sizes.10 to i8*), i32 80, i1 false)
-// CHECK3-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP100:%.*]] = bitcast i8** [[TMP99]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP89]], i32* [[TMP100]], align 4
-// CHECK3-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP89]], i32* [[TMP102]], align 4
-// CHECK3-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP103]], align 4
-// CHECK3-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to [10 x float]**
-// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP105]], align 4
-// CHECK3-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to [10 x float]**
-// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP107]], align 4
-// CHECK3-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP108]], align 4
-// CHECK3-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP110]], align 4
-// CHECK3-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP112]], align 4
-// CHECK3-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP113]], align 4
-// CHECK3-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to float**
-// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP115]], align 4
-// CHECK3-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to float**
-// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP117]], align 4
-// CHECK3-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK3-NEXT:    store i64 [[TMP94]], i64* [[TMP118]], align 4
-// CHECK3-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP103:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    store i32 [[TMP103]], i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK3-NEXT:    [[TMP104:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP104]], i32* [[A_CASTED20]], align 4
+// CHECK3-NEXT:    [[TMP105:%.*]] = load i32, i32* [[A_CASTED20]], align 4
+// CHECK3-NEXT:    [[TMP106:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK3-NEXT:    store i32 [[TMP106]], i32* [[DOTCAPTURE_EXPR__CASTED21]], align 4
+// CHECK3-NEXT:    [[TMP107:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED21]], align 4
+// CHECK3-NEXT:    [[TMP108:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP22:%.*]] = icmp sgt i32 [[TMP108]], 20
+// CHECK3-NEXT:    br i1 [[CMP22]], label [[OMP_IF_THEN23:%.*]], label [[OMP_IF_ELSE31:%.*]]
+// CHECK3:       omp_if.then23:
+// CHECK3-NEXT:    [[TMP109:%.*]] = mul nuw i32 [[TMP1]], 4
+// CHECK3-NEXT:    [[TMP110:%.*]] = sext i32 [[TMP109]] to i64
+// CHECK3-NEXT:    [[TMP111:%.*]] = mul nuw i32 5, [[TMP3]]
+// CHECK3-NEXT:    [[TMP112:%.*]] = mul nuw i32 [[TMP111]], 8
+// CHECK3-NEXT:    [[TMP113:%.*]] = sext i32 [[TMP112]] to i64
+// CHECK3-NEXT:    [[TMP114:%.*]] = bitcast [10 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP114]], i8* align 4 bitcast ([10 x i64]* @.offload_sizes.10 to i8*), i32 80, i1 false)
+// CHECK3-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP105]], i32* [[TMP116]], align 4
+// CHECK3-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP105]], i32* [[TMP118]], align 4
+// CHECK3-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
 // CHECK3-NEXT:    store i8* null, i8** [[TMP119]], align 4
-// CHECK3-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP121:%.*]] = bitcast i8** [[TMP120]] to [5 x [10 x double]]**
-// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP121]], align 4
-// CHECK3-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP123:%.*]] = bitcast i8** [[TMP122]] to [5 x [10 x double]]**
-// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP123]], align 4
-// CHECK3-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP121:%.*]] = bitcast i8** [[TMP120]] to [10 x float]**
+// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP121]], align 4
+// CHECK3-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP123:%.*]] = bitcast i8** [[TMP122]] to [10 x float]**
+// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP123]], align 4
+// CHECK3-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 1
 // CHECK3-NEXT:    store i8* null, i8** [[TMP124]], align 4
-// CHECK3-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i32*
-// CHECK3-NEXT:    store i32 5, i32* [[TMP126]], align 4
-// CHECK3-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 5
+// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP126]], align 4
+// CHECK3-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i32*
-// CHECK3-NEXT:    store i32 5, i32* [[TMP128]], align 4
-// CHECK3-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 5
+// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP128]], align 4
+// CHECK3-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 2
 // CHECK3-NEXT:    store i8* null, i8** [[TMP129]], align 4
-// CHECK3-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 6
-// CHECK3-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP131]], align 4
-// CHECK3-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 6
-// CHECK3-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP133]], align 4
-// CHECK3-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 6
-// CHECK3-NEXT:    store i8* null, i8** [[TMP134]], align 4
-// CHECK3-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 7
-// CHECK3-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to double**
-// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP136]], align 4
-// CHECK3-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 7
-// CHECK3-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to double**
-// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP138]], align 4
-// CHECK3-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK3-NEXT:    store i64 [[TMP97]], i64* [[TMP139]], align 4
-// CHECK3-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to float**
+// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP131]], align 4
+// CHECK3-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to float**
+// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP133]], align 4
+// CHECK3-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK3-NEXT:    store i64 [[TMP110]], i64* [[TMP134]], align 4
+// CHECK3-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 3
+// CHECK3-NEXT:    store i8* null, i8** [[TMP135]], align 4
+// CHECK3-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to [5 x [10 x double]]**
+// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP137]], align 4
+// CHECK3-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to [5 x [10 x double]]**
+// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP139]], align 4
+// CHECK3-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 4
 // CHECK3-NEXT:    store i8* null, i8** [[TMP140]], align 4
-// CHECK3-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 8
-// CHECK3-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to %struct.TT**
-// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP142]], align 4
-// CHECK3-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 8
-// CHECK3-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to %struct.TT**
-// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP144]], align 4
-// CHECK3-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i32*
+// CHECK3-NEXT:    store i32 5, i32* [[TMP142]], align 4
+// CHECK3-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i32*
+// CHECK3-NEXT:    store i32 5, i32* [[TMP144]], align 4
+// CHECK3-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 5
 // CHECK3-NEXT:    store i8* null, i8** [[TMP145]], align 4
-// CHECK3-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 9
+// CHECK3-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 6
 // CHECK3-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP91]], i32* [[TMP147]], align 4
-// CHECK3-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 9
+// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP147]], align 4
+// CHECK3-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 6
 // CHECK3-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP91]], i32* [[TMP149]], align 4
-// CHECK3-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 9
+// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP149]], align 4
+// CHECK3-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 6
 // CHECK3-NEXT:    store i8* null, i8** [[TMP150]], align 4
-// CHECK3-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to double**
+// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP152]], align 4
+// CHECK3-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to double**
+// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP154]], align 4
+// CHECK3-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK3-NEXT:    store i64 [[TMP113]], i64* [[TMP155]], align 4
+// CHECK3-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 7
+// CHECK3-NEXT:    store i8* null, i8** [[TMP156]], align 4
+// CHECK3-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP158:%.*]] = bitcast i8** [[TMP157]] to %struct.TT**
+// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP158]], align 4
+// CHECK3-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP160:%.*]] = bitcast i8** [[TMP159]] to %struct.TT**
+// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP160]], align 4
+// CHECK3-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 8
+// CHECK3-NEXT:    store i8* null, i8** [[TMP161]], align 4
+// CHECK3-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 9
+// CHECK3-NEXT:    [[TMP163:%.*]] = bitcast i8** [[TMP162]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP107]], i32* [[TMP163]], align 4
+// CHECK3-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 9
+// CHECK3-NEXT:    [[TMP165:%.*]] = bitcast i8** [[TMP164]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP107]], i32* [[TMP165]], align 4
+// CHECK3-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 9
+// CHECK3-NEXT:    store i8* null, i8** [[TMP166]], align 4
+// CHECK3-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK3-NEXT:    [[TMP154:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l145.region_id, i32 10, i8** [[TMP151]], i8** [[TMP152]], i64* [[TMP153]], i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.11, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP155:%.*]] = icmp ne i32 [[TMP154]], 0
-// CHECK3-NEXT:    br i1 [[TMP155]], label [[OMP_OFFLOAD_FAILED27:%.*]], label [[OMP_OFFLOAD_CONT28:%.*]]
-// CHECK3:       omp_offload.failed27:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l145(i32 [[TMP89]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP91]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT28]]
-// CHECK3:       omp_offload.cont28:
-// CHECK3-NEXT:    br label [[OMP_IF_END30:%.*]]
-// CHECK3:       omp_if.else29:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l145(i32 [[TMP89]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP91]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_IF_END30]]
-// CHECK3:       omp_if.end30:
-// CHECK3-NEXT:    [[TMP156:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    [[TMP157:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP157]])
-// CHECK3-NEXT:    ret i32 [[TMP156]]
+// CHECK3-NEXT:    [[KERNEL_ARGS28:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP170]], align 4
+// CHECK3-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 10, i32* [[TMP171]], align 4
+// CHECK3-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP167]], i8*** [[TMP172]], align 4
+// CHECK3-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP168]], i8*** [[TMP173]], align 4
+// CHECK3-NEXT:    [[TMP174:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP169]], i64** [[TMP174]], align 4
+// CHECK3-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.11, i32 0, i32 0), i64** [[TMP175]], align 4
+// CHECK3-NEXT:    [[TMP176:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP176]], align 4
+// CHECK3-NEXT:    [[TMP177:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP177]], align 4
+// CHECK3-NEXT:    [[TMP178:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l145.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]])
+// CHECK3-NEXT:    [[TMP179:%.*]] = icmp ne i32 [[TMP178]], 0
+// CHECK3-NEXT:    br i1 [[TMP179]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
+// CHECK3:       omp_offload.failed29:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l145(i32 [[TMP105]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP107]]) #[[ATTR3]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
+// CHECK3:       omp_offload.cont30:
+// CHECK3-NEXT:    br label [[OMP_IF_END32:%.*]]
+// CHECK3:       omp_if.else31:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l145(i32 [[TMP105]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]], i32 [[TMP107]]) #[[ATTR3]]
+// CHECK3-NEXT:    br label [[OMP_IF_END32]]
+// CHECK3:       omp_if.end32:
+// CHECK3-NEXT:    [[TMP180:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    [[TMP181:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP181]])
+// CHECK3-NEXT:    ret i32 [[TMP180]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103
@@ -2445,6 +2615,7 @@ int bar(int n){
 // CHECK3-NEXT:    [[AA_CASTED_I:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED4_I:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK3-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 4
 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -2486,21 +2657,37 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK3-NEXT:    [[TMP26:%.*]] = load i32, i32* [[TMP24]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10) #[[ATTR3]]
-// CHECK3-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 [[TMP26]], i32 0, i8* null, i32 0, i8* null) #[[ATTR3]]
-// CHECK3-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK3-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP27]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP28]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP20]], i8*** [[TMP29]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP30]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP22]], i64** [[TMP31]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP32]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP33]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP34]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP25]], i32 [[TMP26]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR3]]
+// CHECK3-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK3-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK3:       omp_offload.failed.i:
-// CHECK3-NEXT:    [[TMP29:%.*]] = load i16, i16* [[TMP16]], align 2
+// CHECK3-NEXT:    [[TMP37:%.*]] = load i16, i16* [[TMP16]], align 2
 // CHECK3-NEXT:    [[CONV_I:%.*]] = bitcast i32* [[AA_CASTED_I]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP29]], i16* [[CONV_I]], align 2, !noalias !22
-// CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !22
-// CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[TMP23]], align 4
-// CHECK3-NEXT:    store i32 [[TMP31]], i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !22
-// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !22
-// CHECK3-NEXT:    [[TMP33:%.*]] = load i32, i32* [[TMP24]], align 4
-// CHECK3-NEXT:    store i32 [[TMP33]], i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !22
-// CHECK3-NEXT:    [[TMP34:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !22
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103(i32 [[TMP30]], i32 [[TMP32]], i32 [[TMP34]]) #[[ATTR3]]
+// CHECK3-NEXT:    store i16 [[TMP37]], i16* [[CONV_I]], align 2, !noalias !22
+// CHECK3-NEXT:    [[TMP38:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP39:%.*]] = load i32, i32* [[TMP23]], align 4
+// CHECK3-NEXT:    store i32 [[TMP39]], i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP40:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP41:%.*]] = load i32, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    store i32 [[TMP41]], i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !22
+// CHECK3-NEXT:    [[TMP42:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !22
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l103(i32 [[TMP38]], i32 [[TMP40]], i32 [[TMP42]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
 // CHECK3:       .omp_outlined..1.exit:
 // CHECK3-NEXT:    ret i32 0
@@ -3063,9 +3250,26 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK3-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l218.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK3-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.14, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l218.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK3-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l218(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3075,15 +3279,15 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l218(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP42:%.*]] = mul nsw i32 1, [[TMP1]]
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP42]]
+// CHECK3-NEXT:    [[TMP50:%.*]] = mul nsw i32 1, [[TMP1]]
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP50]]
 // CHECK3-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
-// CHECK3-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK3-NEXT:    [[CONV:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK3-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP44]]
-// CHECK3-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK3-NEXT:    [[CONV:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK3-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP52]]
+// CHECK3-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK3-NEXT:    ret i32 [[ADD3]]
 //
 //
@@ -3186,9 +3390,26 @@ int bar(int n){
 // CHECK3-NEXT:    [[ADD6:%.*]] = add i32 [[TMP40]], 1
 // CHECK3-NEXT:    [[TMP41:%.*]] = zext i32 [[ADD6]] to i64
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP41]])
-// CHECK3-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l200.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK3-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP34]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP35]], i8*** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP49]], align 4
+// CHECK3-NEXT:    [[TMP50:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l200.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP51:%.*]] = icmp ne i32 [[TMP50]], 0
+// CHECK3-NEXT:    br i1 [[TMP51]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l200(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], i32 [[TMP7]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3198,8 +3419,8 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l200(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], i32 [[TMP7]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP44:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP44]]
+// CHECK3-NEXT:    [[TMP52:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP52]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -3256,9 +3477,26 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK3-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l183.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.19, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.20, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK3-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.19, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.20, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l183.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK3-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l183(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3268,8 +3506,8 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l183(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP24]]
+// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l218

diff --git a/clang/test/OpenMP/target_teams_distribute_collapse_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_collapse_codegen.cpp
index b08fcec..4962e8e 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_collapse_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_collapse_codegen.cpp
@@ -125,9 +125,26 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 56088)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -135,8 +152,8 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i64 0, i64 0
 // CHECK1-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP9]]
+// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -269,9 +286,26 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 56088)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -279,8 +313,8 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP9]]
+// CHECK3-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -492,20 +526,37 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP46:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_4]], align 8
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP46]], 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK9-NEXT:    [[TMP47:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80.region_id, i32 5, i8** [[TMP39]], i8** [[TMP40]], i64* [[TMP41]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP48:%.*]] = icmp ne i32 [[TMP47]], 0
-// CHECK9-NEXT:    br i1 [[TMP48]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP47]], align 4
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP48]], align 4
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 8
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP40]], i8*** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP41]], i64** [[TMP51]], align 8
+// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP56:%.*]] = icmp ne i32 [[TMP55]], 0
+// CHECK9-NEXT:    br i1 [[TMP56]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80(i64 [[TMP7]], i64 [[TMP9]], i64 [[TMP1]], i64 [[TMP3]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP49:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP49]])
+// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP57]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP50:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP50]])
-// CHECK9-NEXT:    [[TMP51:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP51]]
+// CHECK9-NEXT:    [[TMP58:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP58]])
+// CHECK9-NEXT:    [[TMP59:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP59]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80
@@ -710,9 +761,26 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 20)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l67.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l67.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l67([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -927,20 +995,37 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP45:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_3]], align 8
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP45]], 1
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK11-NEXT:    [[TMP46:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80.region_id, i32 5, i8** [[TMP38]], i8** [[TMP39]], i64* [[TMP40]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP47:%.*]] = icmp ne i32 [[TMP46]], 0
-// CHECK11-NEXT:    br i1 [[TMP47]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP38]], i8*** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP40]], i64** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP55:%.*]] = icmp ne i32 [[TMP54]], 0
+// CHECK11-NEXT:    br i1 [[TMP55]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80(i32 [[TMP5]], i32 [[TMP7]], i32 [[TMP0]], i32 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP48:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP48]])
+// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP56]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP49]])
-// CHECK11-NEXT:    [[TMP50:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP50]]
+// CHECK11-NEXT:    [[TMP57:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP57]])
+// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP58]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80
@@ -1137,9 +1222,26 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 20)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l67.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l67.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l67([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/target_teams_distribute_depend_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_depend_codegen.cpp
index d934048..d0d3f72 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_depend_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_depend_codegen.cpp
@@ -182,7 +182,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 0
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 0, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // CHECK:       [[FAIL]]
@@ -200,7 +200,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 2
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 2, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SZT]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT]], i32 0, i32 0), i8** null, i8** null, i32 0, i32 0, i32 0, i8* null, i32 0, i8* null)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 0, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 
 // CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]

diff --git a/clang/test/OpenMP/target_teams_distribute_dist_schedule_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_dist_schedule_codegen.cpp
index a62dda8..31992d6 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_dist_schedule_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_dist_schedule_codegen.cpp
@@ -142,10 +142,10 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -160,55 +160,106 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK1:       omp_offload.failed16:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK1:       omp_offload.cont15:
-// CHECK1-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP27]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK1:       omp_offload.cont17:
+// CHECK1-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i64 0, i64 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -489,10 +540,10 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -507,55 +558,106 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK3:       omp_offload.failed8:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK3:       omp_offload.cont8:
-// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK3:       omp_offload.cont9:
+// CHECK3-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK3-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK3:       omp_offload.failed14:
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK3:       omp_offload.cont15:
-// CHECK3-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP27]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -838,16 +940,16 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[_TMP9:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
-// CHECK9-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES26:%.*]] = alloca [4 x i64], align 8
+// CHECK9-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_29:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -905,142 +1007,193 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK9-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK9-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK9-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i64 24, i1 false)
-// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK9-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i64 24, i1 false)
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK9-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK9-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK9-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK9-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK9-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK9-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP58]])
-// CHECK9-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK9-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK9:       omp_offload.failed16:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK9:       omp_offload.cont17:
-// CHECK9-NEXT:    [[TMP61:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP61]], i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK9-NEXT:    [[TMP62:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP62]], i32* [[CONV20]], align 4
-// CHECK9-NEXT:    [[TMP63:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK9-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK9-NEXT:    [[CONV21:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP64]], i32* [[CONV21]], align 4
-// CHECK9-NEXT:    [[TMP65:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK9-NEXT:    [[TMP66:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP67:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP67]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.5 to i8*), i64 32, i1 false)
-// CHECK9-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP63]], i64* [[TMP69]], align 8
-// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP63]], i64* [[TMP71]], align 8
-// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP72]], align 8
-// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP74]], align 8
-// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP76]], align 8
-// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP77]], align 8
-// CHECK9-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP79]], align 8
-// CHECK9-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 8
-// CHECK9-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP66]], i64* [[TMP82]], align 8
-// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP83]], align 8
-// CHECK9-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK9-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP66]])
+// CHECK9-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK9-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK9-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK9-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK9-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK9-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK9-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK9:       omp_offload.failed17:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK9:       omp_offload.cont18:
+// CHECK9-NEXT:    [[TMP77:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP77]], i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK9-NEXT:    [[TMP78:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP78]], i32* [[CONV21]], align 4
+// CHECK9-NEXT:    [[TMP79:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK9-NEXT:    [[TMP80:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK9-NEXT:    [[CONV22:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP80]], i32* [[CONV22]], align 4
+// CHECK9-NEXT:    [[TMP81:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK9-NEXT:    [[TMP82:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP83:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES26]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP83]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.5 to i8*), i64 32, i1 false)
+// CHECK9-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP65]], i64* [[TMP85]], align 8
-// CHECK9-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK9-NEXT:    store i64 [[TMP79]], i64* [[TMP85]], align 8
+// CHECK9-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP65]], i64* [[TMP87]], align 8
-// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK9-NEXT:    store i64 [[TMP79]], i64* [[TMP87]], align 8
+// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 0
 // CHECK9-NEXT:    store i8* null, i8** [[TMP88]], align 8
-// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP92:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP92]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK9-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK9-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP93]], 0
-// CHECK9-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
-// CHECK9-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
-// CHECK9-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK9-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK9-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP94]], 1
-// CHECK9-NEXT:    [[TMP95:%.*]] = zext i32 [[ADD32]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP95]])
-// CHECK9-NEXT:    [[TMP96:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, i32 4, i8** [[TMP89]], i8** [[TMP90]], i64* [[TMP91]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP97:%.*]] = icmp ne i32 [[TMP96]], 0
-// CHECK9-NEXT:    br i1 [[TMP97]], label [[OMP_OFFLOAD_FAILED33:%.*]], label [[OMP_OFFLOAD_CONT34:%.*]]
-// CHECK9:       omp_offload.failed33:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i64 [[TMP63]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP65]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT34]]
-// CHECK9:       omp_offload.cont34:
-// CHECK9-NEXT:    [[TMP98:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP98]])
+// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP90]], align 8
+// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP92]], align 8
+// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP93]], align 8
+// CHECK9-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP95]], align 8
+// CHECK9-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 8
+// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP82]], i64* [[TMP98]], align 8
+// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP99]], align 8
+// CHECK9-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP81]], i64* [[TMP101]], align 8
+// CHECK9-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP81]], i64* [[TMP103]], align 8
+// CHECK9-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP104]], align 8
+// CHECK9-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP108:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP108]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK9-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK9-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[TMP109]], 0
+// CHECK9-NEXT:    [[DIV31:%.*]] = sdiv i32 [[SUB30]], 1
+// CHECK9-NEXT:    [[SUB32:%.*]] = sub nsw i32 [[DIV31]], 1
+// CHECK9-NEXT:    store i32 [[SUB32]], i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK9-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK9-NEXT:    [[ADD33:%.*]] = add nsw i32 [[TMP110]], 1
+// CHECK9-NEXT:    [[TMP111:%.*]] = zext i32 [[ADD33]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP111]])
+// CHECK9-NEXT:    [[KERNEL_ARGS34:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP112]], align 4
+// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP113]], align 4
+// CHECK9-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 8
+// CHECK9-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP107]], i64** [[TMP116]], align 8
+// CHECK9-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP117]], align 8
+// CHECK9-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK9-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP119]], align 8
+// CHECK9-NEXT:    [[TMP120:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]])
+// CHECK9-NEXT:    [[TMP121:%.*]] = icmp ne i32 [[TMP120]], 0
+// CHECK9-NEXT:    br i1 [[TMP121]], label [[OMP_OFFLOAD_FAILED35:%.*]], label [[OMP_OFFLOAD_CONT36:%.*]]
+// CHECK9:       omp_offload.failed35:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i64 [[TMP79]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP81]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT36]]
+// CHECK9:       omp_offload.cont36:
+// CHECK9-NEXT:    [[TMP122:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP122]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP99:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP99]])
-// CHECK9-NEXT:    [[TMP100:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP100]]
+// CHECK9-NEXT:    [[TMP123:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP123]])
+// CHECK9-NEXT:    [[TMP124:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP124]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94
@@ -1433,10 +1586,10 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[DOTOFFLOAD_PTRS2:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [1 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [1 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [1 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP1:%.*]] = bitcast i8** [[TMP0]] to [10 x i32]**
@@ -1449,49 +1602,100 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l76.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l76.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l76([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l80.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK9-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK9:       omp_offload.failed5:
+// CHECK9-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK9-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK9-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK9:       omp_offload.failed6:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l80([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK9:       omp_offload.cont6:
-// CHECK9-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP19]], align 8
-// CHECK9-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP21]], align 8
-// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK9:       omp_offload.cont7:
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK9-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK9:       omp_offload.failed11:
+// CHECK9-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK9-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK9-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK9:       omp_offload.failed13:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK9:       omp_offload.cont12:
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK9:       omp_offload.cont14:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -1775,16 +1979,16 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[_TMP8:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_17:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED19:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
-// CHECK11-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES23:%.*]] = alloca [4 x i64], align 4
+// CHECK11-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -1841,141 +2045,192 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK11-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK11-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK11-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK11-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK11-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i32 24, i1 false)
-// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i32 24, i1 false)
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK11-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK11-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK11-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK11-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK11-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK11-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP59]])
-// CHECK11-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK11-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK11:       omp_offload.failed15:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK11:       omp_offload.cont16:
-// CHECK11-NEXT:    [[TMP62:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP62]], i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK11-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP63]], i32* [[N_CASTED18]], align 4
-// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N_CASTED18]], align 4
-// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK11-NEXT:    store i32 [[TMP65]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK11-NEXT:    [[TMP67:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP68:%.*]] = sext i32 [[TMP67]] to i64
-// CHECK11-NEXT:    [[TMP69:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP69]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.5 to i8*), i32 32, i1 false)
-// CHECK11-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[TMP71]], align 4
-// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[TMP73]], align 4
-// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP74]], align 4
-// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP76]], align 4
-// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP78]], align 4
-// CHECK11-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP79]], align 4
-// CHECK11-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 4
-// CHECK11-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 4
-// CHECK11-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP68]], i64* [[TMP84]], align 4
-// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP85]], align 4
-// CHECK11-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK11-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP67]])
+// CHECK11-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK11-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK11-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK11-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK11-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK11-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK11-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK11:       omp_offload.failed16:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK11:       omp_offload.cont17:
+// CHECK11-NEXT:    [[TMP78:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP78]], i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK11-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP79]], i32* [[N_CASTED19]], align 4
+// CHECK11-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_CASTED19]], align 4
+// CHECK11-NEXT:    [[TMP81:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK11-NEXT:    store i32 [[TMP81]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK11-NEXT:    [[TMP82:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK11-NEXT:    [[TMP83:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP84:%.*]] = sext i32 [[TMP83]] to i64
+// CHECK11-NEXT:    [[TMP85:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES23]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP85]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.5 to i8*), i32 32, i1 false)
+// CHECK11-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP66]], i32* [[TMP87]], align 4
-// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK11-NEXT:    store i32 [[TMP80]], i32* [[TMP87]], align 4
+// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP66]], i32* [[TMP89]], align 4
-// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK11-NEXT:    store i32 [[TMP80]], i32* [[TMP89]], align 4
+// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
 // CHECK11-NEXT:    store i8* null, i8** [[TMP90]], align 4
-// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP94:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP94]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP95]], 0
-// CHECK11-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
-// CHECK11-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
-// CHECK11-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK11-NEXT:    [[TMP96:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK11-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP96]], 1
-// CHECK11-NEXT:    [[TMP97:%.*]] = zext i32 [[ADD29]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP97]])
-// CHECK11-NEXT:    [[TMP98:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, i32 4, i8** [[TMP91]], i8** [[TMP92]], i64* [[TMP93]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP99:%.*]] = icmp ne i32 [[TMP98]], 0
-// CHECK11-NEXT:    br i1 [[TMP99]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK11:       omp_offload.failed30:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i32 [[TMP64]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP66]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK11:       omp_offload.cont31:
-// CHECK11-NEXT:    [[TMP100:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP100]])
+// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP92]], align 4
+// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP94]], align 4
+// CHECK11-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP95]], align 4
+// CHECK11-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 4
+// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 4
+// CHECK11-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP84]], i64* [[TMP100]], align 4
+// CHECK11-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP101]], align 4
+// CHECK11-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP82]], i32* [[TMP103]], align 4
+// CHECK11-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP82]], i32* [[TMP105]], align 4
+// CHECK11-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP106]], align 4
+// CHECK11-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP110:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP110]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP111]], 0
+// CHECK11-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
+// CHECK11-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
+// CHECK11-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK11-NEXT:    [[TMP112:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK11-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP112]], 1
+// CHECK11-NEXT:    [[TMP113:%.*]] = zext i32 [[ADD30]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP113]])
+// CHECK11-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP114]], align 4
+// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP115]], align 4
+// CHECK11-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK11-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP108]], i8*** [[TMP117]], align 4
+// CHECK11-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP109]], i64** [[TMP118]], align 4
+// CHECK11-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP119]], align 4
+// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK11-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP121]], align 4
+// CHECK11-NEXT:    [[TMP122:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK11-NEXT:    [[TMP123:%.*]] = icmp ne i32 [[TMP122]], 0
+// CHECK11-NEXT:    br i1 [[TMP123]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK11:       omp_offload.failed32:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i32 [[TMP80]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP82]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK11:       omp_offload.cont33:
+// CHECK11-NEXT:    [[TMP124:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP124]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP101:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP101]])
-// CHECK11-NEXT:    [[TMP102:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP102]]
+// CHECK11-NEXT:    [[TMP125:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP125]])
+// CHECK11-NEXT:    [[TMP126:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP126]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94
@@ -2353,10 +2608,10 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[DOTOFFLOAD_PTRS2:%.*]] = alloca [1 x i8*], align 4
 // CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 4
 // CHECK11-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [1 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [1 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [1 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [1 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [1 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [1 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP1:%.*]] = bitcast i8** [[TMP0]] to [10 x i32]**
@@ -2369,49 +2624,100 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l76.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l76.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l76([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK11-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l80.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK11-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK11:       omp_offload.failed5:
+// CHECK11-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK11-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK11-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK11:       omp_offload.failed6:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l80([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK11:       omp_offload.cont6:
-// CHECK11-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP19]], align 4
-// CHECK11-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP21]], align 4
-// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK11:       omp_offload.cont7:
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK11-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK11:       omp_offload.failed11:
+// CHECK11-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK11-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK11-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK11:       omp_offload.failed13:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK11:       omp_offload.cont12:
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK11:       omp_offload.cont14:
 // CHECK11-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/target_teams_distribute_firstprivate_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_firstprivate_codegen.cpp
index a4babb3..43ed55c 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_firstprivate_codegen.cpp
@@ -314,9 +314,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK1-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK1-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92([2 x i32]* @vec, i64 [[TMP1]], [2 x %struct.S]* @s_arr, %struct.S* @var, i64 [[TMP3]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -588,27 +605,44 @@ int main() {
 // CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK1-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK1-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56([2 x i32]* [[VEC]], i64 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP30]]
+// CHECK1-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN2StC2Ev
@@ -1102,9 +1136,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK3-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK3-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92([2 x i32]* @vec, i32 [[TMP1]], [2 x %struct.S]* @s_arr, %struct.S* @var, i32 [[TMP3]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1367,27 +1418,44 @@ int main() {
 // CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK3-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK3-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56([2 x i32]* [[VEC]], i32 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP30]]
+// CHECK3-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN2StC2Ev

diff --git a/clang/test/OpenMP/target_teams_distribute_lastprivate_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_lastprivate_codegen.cpp
index 8cc392a..7330db6 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_lastprivate_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_lastprivate_codegen.cpp
@@ -563,9 +563,26 @@ int main() {
 // CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 5, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK9-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP35]], align 4
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP32]], i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP33]], i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK9-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94([2 x i32]* [[VEC]], i64 [[TMP3]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -573,18 +590,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP36]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP44]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done3:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP37:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP37]]
+// CHECK9-NEXT:    [[TMP45:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP45]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -877,27 +894,44 @@ int main() {
 // CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK9-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK9-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49([2 x i32]* [[VEC]], i64 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP30]]
+// CHECK9-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1267,9 +1301,26 @@ int main() {
 // CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 5, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK11-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP32]], i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP33]], i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK11-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94([2 x i32]* [[VEC]], i32 [[TMP3]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1277,18 +1328,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP36]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP44]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP37:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP37]]
+// CHECK11-NEXT:    [[TMP45:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP45]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1572,27 +1623,44 @@ int main() {
 // CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK11-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK11-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49([2 x i32]* [[VEC]], i32 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP30]]
+// CHECK11-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_codegen.cpp
index bd88a95..48e7648 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_codegen.cpp
@@ -676,65 +676,99 @@ int target_teams_fun(int *g){
 // CHECK2-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP35]], 1
 // CHECK2-NEXT:    [[TMP36:%.*]] = zext i32 [[ADD]] to i64
 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP36]])
-// CHECK2-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l51.region_id, i32 4, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP31]], i32 [[TMP32]])
-// CHECK2-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
-// CHECK2-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK2-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK2-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK2-NEXT:    store i32 1, i32* [[TMP37]], align 4
+// CHECK2-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK2-NEXT:    store i32 4, i32* [[TMP38]], align 4
+// CHECK2-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK2-NEXT:    store i8** [[TMP29]], i8*** [[TMP39]], align 8
+// CHECK2-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK2-NEXT:    store i8** [[TMP30]], i8*** [[TMP40]], align 8
+// CHECK2-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK2-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP42]], align 8
+// CHECK2-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK2-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP44]], align 8
+// CHECK2-NEXT:    [[TMP45:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 [[TMP31]], i32 [[TMP32]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l51.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK2-NEXT:    [[TMP46:%.*]] = icmp ne i32 [[TMP45]], 0
+// CHECK2-NEXT:    br i1 [[TMP46]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK2:       omp_offload.failed:
 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l51(i64 [[TMP4]], [1000 x i32]* [[A]], i64 [[TMP6]], i64 [[TMP8]]) #[[ATTR2:[0-9]+]]
 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK2:       omp_offload.cont:
-// CHECK2-NEXT:    [[TMP39:%.*]] = load i32, i32* [[N]], align 4
+// CHECK2-NEXT:    [[TMP47:%.*]] = load i32, i32* [[N]], align 4
 // CHECK2-NEXT:    [[CONV10:%.*]] = bitcast i64* [[N_CASTED9]] to i32*
-// CHECK2-NEXT:    store i32 [[TMP39]], i32* [[CONV10]], align 4
-// CHECK2-NEXT:    [[TMP40:%.*]] = load i64, i64* [[N_CASTED9]], align 8
-// CHECK2-NEXT:    [[TMP41:%.*]] = load i32*, i32** [[G_ADDR]], align 8
-// CHECK2-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i64*
-// CHECK2-NEXT:    store i64 [[TMP40]], i64* [[TMP43]], align 8
-// CHECK2-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
-// CHECK2-NEXT:    store i64 [[TMP40]], i64* [[TMP45]], align 8
-// CHECK2-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
-// CHECK2-NEXT:    store i8* null, i8** [[TMP46]], align 8
-// CHECK2-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 1
-// CHECK2-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to [1000 x i32]**
-// CHECK2-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP48]], align 8
-// CHECK2-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 1
-// CHECK2-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to [1000 x i32]**
-// CHECK2-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP50]], align 8
-// CHECK2-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 1
-// CHECK2-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK2-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 2
-// CHECK2-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32**
-// CHECK2-NEXT:    store i32* [[TMP41]], i32** [[TMP53]], align 8
-// CHECK2-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 2
-// CHECK2-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
-// CHECK2-NEXT:    store i32* [[TMP41]], i32** [[TMP55]], align 8
-// CHECK2-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 2
-// CHECK2-NEXT:    store i8* null, i8** [[TMP56]], align 8
-// CHECK2-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP59:%.*]] = load i32, i32* [[N]], align 4
-// CHECK2-NEXT:    store i32 [[TMP59]], i32* [[DOTCAPTURE_EXPR_15]], align 4
-// CHECK2-NEXT:    [[TMP60:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_15]], align 4
-// CHECK2-NEXT:    [[SUB17:%.*]] = sub nsw i32 [[TMP60]], 0
+// CHECK2-NEXT:    store i32 [[TMP47]], i32* [[CONV10]], align 4
+// CHECK2-NEXT:    [[TMP48:%.*]] = load i64, i64* [[N_CASTED9]], align 8
+// CHECK2-NEXT:    [[TMP49:%.*]] = load i32*, i32** [[G_ADDR]], align 8
+// CHECK2-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i64*
+// CHECK2-NEXT:    store i64 [[TMP48]], i64* [[TMP51]], align 8
+// CHECK2-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i64*
+// CHECK2-NEXT:    store i64 [[TMP48]], i64* [[TMP53]], align 8
+// CHECK2-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK2-NEXT:    store i8* null, i8** [[TMP54]], align 8
+// CHECK2-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 1
+// CHECK2-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to [1000 x i32]**
+// CHECK2-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP56]], align 8
+// CHECK2-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 1
+// CHECK2-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to [1000 x i32]**
+// CHECK2-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP58]], align 8
+// CHECK2-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 1
+// CHECK2-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK2-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 2
+// CHECK2-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to i32**
+// CHECK2-NEXT:    store i32* [[TMP49]], i32** [[TMP61]], align 8
+// CHECK2-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 2
+// CHECK2-NEXT:    [[TMP63:%.*]] = bitcast i8** [[TMP62]] to i32**
+// CHECK2-NEXT:    store i32* [[TMP49]], i32** [[TMP63]], align 8
+// CHECK2-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 2
+// CHECK2-NEXT:    store i8* null, i8** [[TMP64]], align 8
+// CHECK2-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP67:%.*]] = load i32, i32* [[N]], align 4
+// CHECK2-NEXT:    store i32 [[TMP67]], i32* [[DOTCAPTURE_EXPR_15]], align 4
+// CHECK2-NEXT:    [[TMP68:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_15]], align 4
+// CHECK2-NEXT:    [[SUB17:%.*]] = sub nsw i32 [[TMP68]], 0
 // CHECK2-NEXT:    [[DIV18:%.*]] = sdiv i32 [[SUB17]], 1
 // CHECK2-NEXT:    [[SUB19:%.*]] = sub nsw i32 [[DIV18]], 1
 // CHECK2-NEXT:    store i32 [[SUB19]], i32* [[DOTCAPTURE_EXPR_16]], align 4
-// CHECK2-NEXT:    [[TMP61:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_16]], align 4
-// CHECK2-NEXT:    [[ADD20:%.*]] = add nsw i32 [[TMP61]], 1
-// CHECK2-NEXT:    [[TMP62:%.*]] = zext i32 [[ADD20]] to i64
-// CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP62]])
-// CHECK2-NEXT:    [[TMP63:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l58.region_id, i32 3, i8** [[TMP57]], i8** [[TMP58]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK2-NEXT:    [[TMP64:%.*]] = icmp ne i32 [[TMP63]], 0
-// CHECK2-NEXT:    br i1 [[TMP64]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK2:       omp_offload.failed21:
-// CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l58(i64 [[TMP40]], [1000 x i32]* [[A]], i32* [[TMP41]]) #[[ATTR2]]
-// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK2:       omp_offload.cont22:
+// CHECK2-NEXT:    [[TMP69:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_16]], align 4
+// CHECK2-NEXT:    [[ADD20:%.*]] = add nsw i32 [[TMP69]], 1
+// CHECK2-NEXT:    [[TMP70:%.*]] = zext i32 [[ADD20]] to i64
+// CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP70]])
+// CHECK2-NEXT:    [[KERNEL_ARGS21:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK2-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS21]], i32 0, i32 0
+// CHECK2-NEXT:    store i32 1, i32* [[TMP71]], align 4
+// CHECK2-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS21]], i32 0, i32 1
+// CHECK2-NEXT:    store i32 3, i32* [[TMP72]], align 4
+// CHECK2-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS21]], i32 0, i32 2
+// CHECK2-NEXT:    store i8** [[TMP65]], i8*** [[TMP73]], align 8
+// CHECK2-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS21]], i32 0, i32 3
+// CHECK2-NEXT:    store i8** [[TMP66]], i8*** [[TMP74]], align 8
+// CHECK2-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS21]], i32 0, i32 4
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP75]], align 8
+// CHECK2-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS21]], i32 0, i32 5
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP76]], align 8
+// CHECK2-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS21]], i32 0, i32 6
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP77]], align 8
+// CHECK2-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS21]], i32 0, i32 7
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP78]], align 8
+// CHECK2-NEXT:    [[TMP79:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l58.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS21]])
+// CHECK2-NEXT:    [[TMP80:%.*]] = icmp ne i32 [[TMP79]], 0
+// CHECK2-NEXT:    br i1 [[TMP80]], label [[OMP_OFFLOAD_FAILED22:%.*]], label [[OMP_OFFLOAD_CONT23:%.*]]
+// CHECK2:       omp_offload.failed22:
+// CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l58(i64 [[TMP48]], [1000 x i32]* [[A]], i32* [[TMP49]]) #[[ATTR2]]
+// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT23]]
+// CHECK2:       omp_offload.cont23:
 // CHECK2-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [1000 x i32], [1000 x i32]* [[A]], i64 0, i64 0
-// CHECK2-NEXT:    [[TMP65:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK2-NEXT:    ret i32 [[TMP65]]
+// CHECK2-NEXT:    [[TMP81:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK2-NEXT:    ret i32 [[TMP81]]
 //
 //
 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l51
@@ -1298,64 +1332,98 @@ int target_teams_fun(int *g){
 // CHECK4-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP35]], 1
 // CHECK4-NEXT:    [[TMP36:%.*]] = zext i32 [[ADD]] to i64
 // CHECK4-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP36]])
-// CHECK4-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l51.region_id, i32 4, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP31]], i32 [[TMP32]])
-// CHECK4-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
-// CHECK4-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK4-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK4-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK4-NEXT:    store i32 1, i32* [[TMP37]], align 4
+// CHECK4-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK4-NEXT:    store i32 4, i32* [[TMP38]], align 4
+// CHECK4-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK4-NEXT:    store i8** [[TMP29]], i8*** [[TMP39]], align 4
+// CHECK4-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK4-NEXT:    store i8** [[TMP30]], i8*** [[TMP40]], align 4
+// CHECK4-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK4-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK4-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK4-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP42]], align 4
+// CHECK4-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK4-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK4-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK4-NEXT:    store i8** null, i8*** [[TMP44]], align 4
+// CHECK4-NEXT:    [[TMP45:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 [[TMP31]], i32 [[TMP32]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l51.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK4-NEXT:    [[TMP46:%.*]] = icmp ne i32 [[TMP45]], 0
+// CHECK4-NEXT:    br i1 [[TMP46]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK4:       omp_offload.failed:
 // CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l51(i32 [[TMP4]], [1000 x i32]* [[A]], i32 [[TMP6]], i32 [[TMP8]]) #[[ATTR2:[0-9]+]]
 // CHECK4-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK4:       omp_offload.cont:
-// CHECK4-NEXT:    [[TMP39:%.*]] = load i32, i32* [[N]], align 4
-// CHECK4-NEXT:    store i32 [[TMP39]], i32* [[N_CASTED7]], align 4
-// CHECK4-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N_CASTED7]], align 4
-// CHECK4-NEXT:    [[TMP41:%.*]] = load i32*, i32** [[G_ADDR]], align 4
-// CHECK4-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
-// CHECK4-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK4-NEXT:    store i32 [[TMP40]], i32* [[TMP43]], align 4
-// CHECK4-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
-// CHECK4-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK4-NEXT:    store i32 [[TMP40]], i32* [[TMP45]], align 4
-// CHECK4-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
-// CHECK4-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK4-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
-// CHECK4-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to [1000 x i32]**
-// CHECK4-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP48]], align 4
-// CHECK4-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
-// CHECK4-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to [1000 x i32]**
-// CHECK4-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP50]], align 4
-// CHECK4-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
-// CHECK4-NEXT:    store i8* null, i8** [[TMP51]], align 4
-// CHECK4-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 2
-// CHECK4-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32**
-// CHECK4-NEXT:    store i32* [[TMP41]], i32** [[TMP53]], align 4
-// CHECK4-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 2
-// CHECK4-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
-// CHECK4-NEXT:    store i32* [[TMP41]], i32** [[TMP55]], align 4
-// CHECK4-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 2
-// CHECK4-NEXT:    store i8* null, i8** [[TMP56]], align 4
-// CHECK4-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
-// CHECK4-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
-// CHECK4-NEXT:    [[TMP59:%.*]] = load i32, i32* [[N]], align 4
-// CHECK4-NEXT:    store i32 [[TMP59]], i32* [[DOTCAPTURE_EXPR_12]], align 4
-// CHECK4-NEXT:    [[TMP60:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_12]], align 4
-// CHECK4-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[TMP60]], 0
+// CHECK4-NEXT:    [[TMP47:%.*]] = load i32, i32* [[N]], align 4
+// CHECK4-NEXT:    store i32 [[TMP47]], i32* [[N_CASTED7]], align 4
+// CHECK4-NEXT:    [[TMP48:%.*]] = load i32, i32* [[N_CASTED7]], align 4
+// CHECK4-NEXT:    [[TMP49:%.*]] = load i32*, i32** [[G_ADDR]], align 4
+// CHECK4-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK4-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK4-NEXT:    store i32 [[TMP48]], i32* [[TMP51]], align 4
+// CHECK4-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK4-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK4-NEXT:    store i32 [[TMP48]], i32* [[TMP53]], align 4
+// CHECK4-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK4-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK4-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK4-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to [1000 x i32]**
+// CHECK4-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP56]], align 4
+// CHECK4-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK4-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to [1000 x i32]**
+// CHECK4-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP58]], align 4
+// CHECK4-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK4-NEXT:    store i8* null, i8** [[TMP59]], align 4
+// CHECK4-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 2
+// CHECK4-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to i32**
+// CHECK4-NEXT:    store i32* [[TMP49]], i32** [[TMP61]], align 4
+// CHECK4-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 2
+// CHECK4-NEXT:    [[TMP63:%.*]] = bitcast i8** [[TMP62]] to i32**
+// CHECK4-NEXT:    store i32* [[TMP49]], i32** [[TMP63]], align 4
+// CHECK4-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 2
+// CHECK4-NEXT:    store i8* null, i8** [[TMP64]], align 4
+// CHECK4-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK4-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK4-NEXT:    [[TMP67:%.*]] = load i32, i32* [[N]], align 4
+// CHECK4-NEXT:    store i32 [[TMP67]], i32* [[DOTCAPTURE_EXPR_12]], align 4
+// CHECK4-NEXT:    [[TMP68:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_12]], align 4
+// CHECK4-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[TMP68]], 0
 // CHECK4-NEXT:    [[DIV15:%.*]] = sdiv i32 [[SUB14]], 1
 // CHECK4-NEXT:    [[SUB16:%.*]] = sub nsw i32 [[DIV15]], 1
 // CHECK4-NEXT:    store i32 [[SUB16]], i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK4-NEXT:    [[TMP61:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK4-NEXT:    [[ADD17:%.*]] = add nsw i32 [[TMP61]], 1
-// CHECK4-NEXT:    [[TMP62:%.*]] = zext i32 [[ADD17]] to i64
-// CHECK4-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP62]])
-// CHECK4-NEXT:    [[TMP63:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l58.region_id, i32 3, i8** [[TMP57]], i8** [[TMP58]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK4-NEXT:    [[TMP64:%.*]] = icmp ne i32 [[TMP63]], 0
-// CHECK4-NEXT:    br i1 [[TMP64]], label [[OMP_OFFLOAD_FAILED18:%.*]], label [[OMP_OFFLOAD_CONT19:%.*]]
-// CHECK4:       omp_offload.failed18:
-// CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l58(i32 [[TMP40]], [1000 x i32]* [[A]], i32* [[TMP41]]) #[[ATTR2]]
-// CHECK4-NEXT:    br label [[OMP_OFFLOAD_CONT19]]
-// CHECK4:       omp_offload.cont19:
+// CHECK4-NEXT:    [[TMP69:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
+// CHECK4-NEXT:    [[ADD17:%.*]] = add nsw i32 [[TMP69]], 1
+// CHECK4-NEXT:    [[TMP70:%.*]] = zext i32 [[ADD17]] to i64
+// CHECK4-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP70]])
+// CHECK4-NEXT:    [[KERNEL_ARGS18:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK4-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 0
+// CHECK4-NEXT:    store i32 1, i32* [[TMP71]], align 4
+// CHECK4-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 1
+// CHECK4-NEXT:    store i32 3, i32* [[TMP72]], align 4
+// CHECK4-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 2
+// CHECK4-NEXT:    store i8** [[TMP65]], i8*** [[TMP73]], align 4
+// CHECK4-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 3
+// CHECK4-NEXT:    store i8** [[TMP66]], i8*** [[TMP74]], align 4
+// CHECK4-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 4
+// CHECK4-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP75]], align 4
+// CHECK4-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 5
+// CHECK4-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP76]], align 4
+// CHECK4-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 6
+// CHECK4-NEXT:    store i8** null, i8*** [[TMP77]], align 4
+// CHECK4-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 7
+// CHECK4-NEXT:    store i8** null, i8*** [[TMP78]], align 4
+// CHECK4-NEXT:    [[TMP79:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l58.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]])
+// CHECK4-NEXT:    [[TMP80:%.*]] = icmp ne i32 [[TMP79]], 0
+// CHECK4-NEXT:    br i1 [[TMP80]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
+// CHECK4:       omp_offload.failed19:
+// CHECK4-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l58(i32 [[TMP48]], [1000 x i32]* [[A]], i32* [[TMP49]]) #[[ATTR2]]
+// CHECK4-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
+// CHECK4:       omp_offload.cont20:
 // CHECK4-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [1000 x i32], [1000 x i32]* [[A]], i32 0, i32 0
-// CHECK4-NEXT:    [[TMP65:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK4-NEXT:    ret i32 [[TMP65]]
+// CHECK4-NEXT:    [[TMP81:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK4-NEXT:    ret i32 [[TMP81]]
 //
 //
 // CHECK4-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l51

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_collapse_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_collapse_codegen.cpp
index 8ef7cb7..5cf898e 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_collapse_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_collapse_codegen.cpp
@@ -130,9 +130,26 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 56088)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -140,8 +157,8 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i64 0, i64 0
 // CHECK1-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP9]]
+// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -349,9 +366,26 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 56088)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -359,8 +393,8 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP9]]
+// CHECK3-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -643,20 +677,37 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP46:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_4]], align 8
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP46]], 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK9-NEXT:    [[TMP47:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, i32 5, i8** [[TMP39]], i8** [[TMP40]], i64* [[TMP41]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP48:%.*]] = icmp ne i32 [[TMP47]], 0
-// CHECK9-NEXT:    br i1 [[TMP48]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP47]], align 4
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP48]], align 4
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 8
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP40]], i8*** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP41]], i64** [[TMP51]], align 8
+// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP56:%.*]] = icmp ne i32 [[TMP55]], 0
+// CHECK9-NEXT:    br i1 [[TMP56]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81(i64 [[TMP7]], i64 [[TMP9]], i64 [[TMP1]], i64 [[TMP3]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP49:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP49]])
+// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP57]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP50:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP50]])
-// CHECK9-NEXT:    [[TMP51:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP51]]
+// CHECK9-NEXT:    [[TMP58:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP58]])
+// CHECK9-NEXT:    [[TMP59:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP59]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81
@@ -991,9 +1042,26 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 20)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1283,20 +1351,37 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP45:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_3]], align 8
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP45]], 1
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK11-NEXT:    [[TMP46:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, i32 5, i8** [[TMP38]], i8** [[TMP39]], i64* [[TMP40]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP47:%.*]] = icmp ne i32 [[TMP46]], 0
-// CHECK11-NEXT:    br i1 [[TMP47]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP38]], i8*** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP40]], i64** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP55:%.*]] = icmp ne i32 [[TMP54]], 0
+// CHECK11-NEXT:    br i1 [[TMP55]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81(i32 [[TMP5]], i32 [[TMP7]], i32 [[TMP0]], i32 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP48:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP48]])
+// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP56]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP49]])
-// CHECK11-NEXT:    [[TMP50:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP50]]
+// CHECK11-NEXT:    [[TMP57:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP57]])
+// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP58]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81
@@ -1623,9 +1708,26 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 20)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_depend_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_depend_codegen.cpp
index e2bda82..67fcde3 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_depend_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_depend_codegen.cpp
@@ -182,7 +182,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 0
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 0, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // CHECK:       [[FAIL]]
@@ -200,7 +200,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 2
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 2, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SZT]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT]], i32 0, i32 0), i8** null, i8** null, i32 0, i32 0, i32 0, i8* null, i32 0, i8* null)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 0, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 
 // CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_dist_schedule_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_dist_schedule_codegen.cpp
index aa78061..60701e1 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_dist_schedule_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_dist_schedule_codegen.cpp
@@ -154,10 +154,10 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -172,55 +172,106 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK1:       omp_offload.failed16:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK1:       omp_offload.cont15:
-// CHECK1-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP27]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK1:       omp_offload.cont17:
+// CHECK1-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i64 0, i64 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -723,10 +774,10 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -741,55 +792,106 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK3:       omp_offload.failed8:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK3:       omp_offload.cont8:
-// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK3:       omp_offload.cont9:
+// CHECK3-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK3-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK3:       omp_offload.failed14:
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK3:       omp_offload.cont15:
-// CHECK3-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP27]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -1283,16 +1385,16 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[_TMP9:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
-// CHECK9-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES26:%.*]] = alloca [4 x i64], align 8
+// CHECK9-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_29:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -1351,142 +1453,193 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK9-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK9-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK9-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
-// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK9-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK9-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK9-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK9-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK9-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK9-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK9-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP58]])
-// CHECK9-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK9-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK9:       omp_offload.failed16:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK9:       omp_offload.cont17:
-// CHECK9-NEXT:    [[TMP61:%.*]] = load i32, i32* [[M]], align 4
-// CHECK9-NEXT:    store i32 [[TMP61]], i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK9-NEXT:    [[TMP62:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP62]], i32* [[CONV20]], align 4
-// CHECK9-NEXT:    [[TMP63:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK9-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK9-NEXT:    [[CONV21:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP64]], i32* [[CONV21]], align 4
-// CHECK9-NEXT:    [[TMP65:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK9-NEXT:    [[TMP66:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP67:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP67]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
-// CHECK9-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP63]], i64* [[TMP69]], align 8
-// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP63]], i64* [[TMP71]], align 8
-// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP72]], align 8
-// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP74]], align 8
-// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP76]], align 8
-// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP77]], align 8
-// CHECK9-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP79]], align 8
-// CHECK9-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 8
-// CHECK9-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP66]], i64* [[TMP82]], align 8
-// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP83]], align 8
-// CHECK9-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK9-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP66]])
+// CHECK9-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK9-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK9-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK9-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK9-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK9-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK9-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK9:       omp_offload.failed17:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK9:       omp_offload.cont18:
+// CHECK9-NEXT:    [[TMP77:%.*]] = load i32, i32* [[M]], align 4
+// CHECK9-NEXT:    store i32 [[TMP77]], i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK9-NEXT:    [[TMP78:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP78]], i32* [[CONV21]], align 4
+// CHECK9-NEXT:    [[TMP79:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK9-NEXT:    [[TMP80:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK9-NEXT:    [[CONV22:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP80]], i32* [[CONV22]], align 4
+// CHECK9-NEXT:    [[TMP81:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK9-NEXT:    [[TMP82:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP83:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES26]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP83]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
+// CHECK9-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP65]], i64* [[TMP85]], align 8
-// CHECK9-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK9-NEXT:    store i64 [[TMP79]], i64* [[TMP85]], align 8
+// CHECK9-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP65]], i64* [[TMP87]], align 8
-// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK9-NEXT:    store i64 [[TMP79]], i64* [[TMP87]], align 8
+// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 0
 // CHECK9-NEXT:    store i8* null, i8** [[TMP88]], align 8
-// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP92:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP92]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK9-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK9-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP93]], 0
-// CHECK9-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
-// CHECK9-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
-// CHECK9-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK9-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK9-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP94]], 1
-// CHECK9-NEXT:    [[TMP95:%.*]] = zext i32 [[ADD32]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP95]])
-// CHECK9-NEXT:    [[TMP96:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108.region_id, i32 4, i8** [[TMP89]], i8** [[TMP90]], i64* [[TMP91]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP97:%.*]] = icmp ne i32 [[TMP96]], 0
-// CHECK9-NEXT:    br i1 [[TMP97]], label [[OMP_OFFLOAD_FAILED33:%.*]], label [[OMP_OFFLOAD_CONT34:%.*]]
-// CHECK9:       omp_offload.failed33:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108(i64 [[TMP63]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP65]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT34]]
-// CHECK9:       omp_offload.cont34:
-// CHECK9-NEXT:    [[TMP98:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP98]])
+// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP90]], align 8
+// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP92]], align 8
+// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP93]], align 8
+// CHECK9-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP95]], align 8
+// CHECK9-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 8
+// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP82]], i64* [[TMP98]], align 8
+// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP99]], align 8
+// CHECK9-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP81]], i64* [[TMP101]], align 8
+// CHECK9-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP81]], i64* [[TMP103]], align 8
+// CHECK9-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP104]], align 8
+// CHECK9-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP108:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP108]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK9-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK9-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[TMP109]], 0
+// CHECK9-NEXT:    [[DIV31:%.*]] = sdiv i32 [[SUB30]], 1
+// CHECK9-NEXT:    [[SUB32:%.*]] = sub nsw i32 [[DIV31]], 1
+// CHECK9-NEXT:    store i32 [[SUB32]], i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK9-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK9-NEXT:    [[ADD33:%.*]] = add nsw i32 [[TMP110]], 1
+// CHECK9-NEXT:    [[TMP111:%.*]] = zext i32 [[ADD33]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP111]])
+// CHECK9-NEXT:    [[KERNEL_ARGS34:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP112]], align 4
+// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP113]], align 4
+// CHECK9-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 8
+// CHECK9-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP107]], i64** [[TMP116]], align 8
+// CHECK9-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP117]], align 8
+// CHECK9-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK9-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP119]], align 8
+// CHECK9-NEXT:    [[TMP120:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]])
+// CHECK9-NEXT:    [[TMP121:%.*]] = icmp ne i32 [[TMP120]], 0
+// CHECK9-NEXT:    br i1 [[TMP121]], label [[OMP_OFFLOAD_FAILED35:%.*]], label [[OMP_OFFLOAD_CONT36:%.*]]
+// CHECK9:       omp_offload.failed35:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108(i64 [[TMP79]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP81]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT36]]
+// CHECK9:       omp_offload.cont36:
+// CHECK9-NEXT:    [[TMP122:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP122]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP99:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP99]])
-// CHECK9-NEXT:    [[TMP100:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP100]]
+// CHECK9-NEXT:    [[TMP123:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP123]])
+// CHECK9-NEXT:    [[TMP124:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP124]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100
@@ -2215,10 +2368,10 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -2232,63 +2385,114 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l81.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l81([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l85.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK9-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK9:       omp_offload.failed5:
+// CHECK9-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l85.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK9-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK9-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK9:       omp_offload.failed6:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l85([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK9:       omp_offload.cont6:
-// CHECK9-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK9-NEXT:    store i32 [[TMP18]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK9-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK9:       omp_offload.cont7:
+// CHECK9-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK9-NEXT:    store i32 [[TMP34]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK9-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP19]], i32* [[CONV]], align 4
-// CHECK9-NEXT:    [[TMP20:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP22]], align 8
-// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP24]], align 8
-// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP25]], align 8
-// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP20]], i64* [[TMP27]], align 8
-// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP20]], i64* [[TMP29]], align 8
-// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP30]], align 8
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 [[TMP35]], i32* [[CONV]], align 4
+// CHECK9-NEXT:    [[TMP36:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP36]], i64* [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP36]], i64* [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP46]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, i32 2, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK9-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK9:       omp_offload.failed11:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]], i64 [[TMP20]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK9:       omp_offload.cont12:
+// CHECK9-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP49]], align 4
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 2, i32* [[TMP50]], align 4
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 8
+// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP48]], i8*** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP56]], align 8
+// CHECK9-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK9-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
+// CHECK9-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK9:       omp_offload.failed13:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]], i64 [[TMP36]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK9:       omp_offload.cont14:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -2815,16 +3019,16 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[_TMP8:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_17:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED19:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
-// CHECK11-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES23:%.*]] = alloca [4 x i64], align 4
+// CHECK11-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -2882,141 +3086,192 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK11-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK11-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK11-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK11-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK11-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
-// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK11-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK11-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK11-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK11-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK11-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK11-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP59]])
-// CHECK11-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK11-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK11:       omp_offload.failed15:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK11:       omp_offload.cont16:
-// CHECK11-NEXT:    [[TMP62:%.*]] = load i32, i32* [[M]], align 4
-// CHECK11-NEXT:    store i32 [[TMP62]], i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK11-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP63]], i32* [[N_CASTED18]], align 4
-// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N_CASTED18]], align 4
-// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK11-NEXT:    store i32 [[TMP65]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK11-NEXT:    [[TMP67:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP68:%.*]] = sext i32 [[TMP67]] to i64
-// CHECK11-NEXT:    [[TMP69:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP69]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
-// CHECK11-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[TMP71]], align 4
-// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[TMP73]], align 4
-// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP74]], align 4
-// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP76]], align 4
-// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP78]], align 4
-// CHECK11-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP79]], align 4
-// CHECK11-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 4
-// CHECK11-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 4
-// CHECK11-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP68]], i64* [[TMP84]], align 4
-// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP85]], align 4
-// CHECK11-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK11-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP67]])
+// CHECK11-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK11-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK11-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK11-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK11-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK11-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK11-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK11:       omp_offload.failed16:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK11:       omp_offload.cont17:
+// CHECK11-NEXT:    [[TMP78:%.*]] = load i32, i32* [[M]], align 4
+// CHECK11-NEXT:    store i32 [[TMP78]], i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK11-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP79]], i32* [[N_CASTED19]], align 4
+// CHECK11-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_CASTED19]], align 4
+// CHECK11-NEXT:    [[TMP81:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK11-NEXT:    store i32 [[TMP81]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK11-NEXT:    [[TMP82:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK11-NEXT:    [[TMP83:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP84:%.*]] = sext i32 [[TMP83]] to i64
+// CHECK11-NEXT:    [[TMP85:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES23]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP85]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
+// CHECK11-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP66]], i32* [[TMP87]], align 4
-// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK11-NEXT:    store i32 [[TMP80]], i32* [[TMP87]], align 4
+// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP66]], i32* [[TMP89]], align 4
-// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK11-NEXT:    store i32 [[TMP80]], i32* [[TMP89]], align 4
+// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
 // CHECK11-NEXT:    store i8* null, i8** [[TMP90]], align 4
-// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP94:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP94]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP95]], 0
-// CHECK11-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
-// CHECK11-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
-// CHECK11-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK11-NEXT:    [[TMP96:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK11-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP96]], 1
-// CHECK11-NEXT:    [[TMP97:%.*]] = zext i32 [[ADD29]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP97]])
-// CHECK11-NEXT:    [[TMP98:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108.region_id, i32 4, i8** [[TMP91]], i8** [[TMP92]], i64* [[TMP93]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP99:%.*]] = icmp ne i32 [[TMP98]], 0
-// CHECK11-NEXT:    br i1 [[TMP99]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK11:       omp_offload.failed30:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108(i32 [[TMP64]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP66]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK11:       omp_offload.cont31:
-// CHECK11-NEXT:    [[TMP100:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP100]])
+// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP92]], align 4
+// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP94]], align 4
+// CHECK11-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP95]], align 4
+// CHECK11-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 4
+// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 4
+// CHECK11-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP84]], i64* [[TMP100]], align 4
+// CHECK11-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP101]], align 4
+// CHECK11-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP82]], i32* [[TMP103]], align 4
+// CHECK11-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP82]], i32* [[TMP105]], align 4
+// CHECK11-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP106]], align 4
+// CHECK11-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP110:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP110]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP111]], 0
+// CHECK11-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
+// CHECK11-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
+// CHECK11-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK11-NEXT:    [[TMP112:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK11-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP112]], 1
+// CHECK11-NEXT:    [[TMP113:%.*]] = zext i32 [[ADD30]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP113]])
+// CHECK11-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP114]], align 4
+// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP115]], align 4
+// CHECK11-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK11-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP108]], i8*** [[TMP117]], align 4
+// CHECK11-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP109]], i64** [[TMP118]], align 4
+// CHECK11-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP119]], align 4
+// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK11-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP121]], align 4
+// CHECK11-NEXT:    [[TMP122:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK11-NEXT:    [[TMP123:%.*]] = icmp ne i32 [[TMP122]], 0
+// CHECK11-NEXT:    br i1 [[TMP123]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK11:       omp_offload.failed32:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108(i32 [[TMP80]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP82]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK11:       omp_offload.cont33:
+// CHECK11-NEXT:    [[TMP124:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP124]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP101:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP101]])
-// CHECK11-NEXT:    [[TMP102:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP102]]
+// CHECK11-NEXT:    [[TMP125:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP125]])
+// CHECK11-NEXT:    [[TMP126:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP126]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100
@@ -3710,10 +3965,10 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -3727,62 +3982,113 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l81.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l81([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK11-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l85.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK11-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK11:       omp_offload.failed5:
+// CHECK11-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l85.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK11-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK11-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK11:       omp_offload.failed6:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l85([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK11:       omp_offload.cont6:
-// CHECK11-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK11-NEXT:    store i32 [[TMP18]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK11-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK11-NEXT:    store i32 [[TMP19]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK11-NEXT:    [[TMP20:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP22]], align 4
-// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP24]], align 4
-// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP25]], align 4
-// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP20]], i32* [[TMP27]], align 4
-// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP20]], i32* [[TMP29]], align 4
-// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP30]], align 4
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK11:       omp_offload.cont7:
+// CHECK11-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK11-NEXT:    store i32 [[TMP34]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK11-NEXT:    store i32 [[TMP35]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP36]], i32* [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP36]], i32* [[TMP45]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, i32 2, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK11-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK11:       omp_offload.failed11:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]], i32 [[TMP20]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK11:       omp_offload.cont12:
+// CHECK11-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 2, i32* [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP48]], i8*** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP55]], align 4
+// CHECK11-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP56]], align 4
+// CHECK11-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK11-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
+// CHECK11-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK11:       omp_offload.failed13:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]], i32 [[TMP36]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK11:       omp_offload.cont14:
 // CHECK11-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_firstprivate_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_firstprivate_codegen.cpp
index 16f2cd0..c42e7f1 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_firstprivate_codegen.cpp
@@ -375,9 +375,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l122.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK1-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l122.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK1-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l122([2 x i32]* @vec, i64 [[TMP1]], [2 x %struct.S]* @s_arr, %struct.S* @var, i64 [[TMP3]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -786,27 +803,44 @@ int main() {
 // CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK1-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i64** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK1-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81([2 x i32]* [[VEC]], i64 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP30]]
+// CHECK1-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN2StC2Ev
@@ -1435,9 +1469,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l122.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK3-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l122.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK3-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l122([2 x i32]* @vec, i32 [[TMP1]], [2 x %struct.S]* @s_arr, %struct.S* @var, i32 [[TMP3]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1829,27 +1880,44 @@ int main() {
 // CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK3-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i64** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK3-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81([2 x i32]* [[VEC]], i32 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP30]]
+// CHECK3-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN2StC2Ev

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_if_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_if_codegen.cpp
index e993283..8cfbb53 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_if_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_if_codegen.cpp
@@ -111,21 +111,55 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48() #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l51.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l51.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l51() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret void
 //
 //
@@ -422,56 +456,90 @@ int main() {
 // CHECK1-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83() #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* @Arg, align 4
-// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP2]], 0
+// CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP10]], 0
 // CHECK1-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP3:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP3]] to i1
+// CHECK1-NEXT:    [[TMP11:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP11]] to i1
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i8*
 // CHECK1-NEXT:    [[FROMBOOL2:%.*]] = zext i1 [[TOBOOL1]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL2]], i8* [[CONV]], align 1
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP5]] to i1
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP13]] to i1
 // CHECK1-NEXT:    br i1 [[TOBOOL3]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK1:       omp_if.then:
-// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP7]], align 8
-// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP9]], align 8
-// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP13:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP13]] to i1
-// CHECK1-NEXT:    [[TMP14:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP21]] to i1
+// CHECK1-NEXT:    [[TMP22:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP14]])
-// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
-// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK1:       omp_offload.failed6:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90(i64 [[TMP4]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK1:       omp_offload.cont7:
+// CHECK1-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP19]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP22]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK1-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
+// CHECK1-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90(i64 [[TMP12]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK1:       omp_offload.cont8:
 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK1:       omp_if.else:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90(i64 [[TMP4]]) #[[ATTR2]]
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l90(i64 [[TMP12]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* @Arg, align 4
-// CHECK1-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP17]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP33]])
 // CHECK1-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -934,45 +1002,79 @@ int main() {
 // CHECK1-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l60.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l60.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l60() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l64() #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
-// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP2]], 0
+// CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP10]], 0
 // CHECK1-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP3:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP3]] to i1
+// CHECK1-NEXT:    [[TMP11:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP11]] to i1
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i8*
 // CHECK1-NEXT:    [[FROMBOOL2:%.*]] = zext i1 [[TOBOOL1]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL2]], i8* [[CONV]], align 1
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP6]], align 8
-// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP8]], align 8
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP9]], align 8
-// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP12]] to i1
-// CHECK1-NEXT:    [[TMP13:%.*]] = select i1 [[TOBOOL3]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP14:%.*]] = bitcast i8** [[TMP13]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP20]] to i1
+// CHECK1-NEXT:    [[TMP21:%.*]] = select i1 [[TOBOOL3]], i32 0, i32 1
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l68.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP13]])
-// CHECK1-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK1-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK1:       omp_offload.failed5:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l68(i64 [[TMP4]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK1:       omp_offload.cont6:
+// CHECK1-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP18]], i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP19]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP21]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK1-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK1-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK1:       omp_offload.failed6:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l68(i64 [[TMP12]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK1:       omp_offload.cont7:
 // CHECK1-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_lastprivate_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_lastprivate_codegen.cpp
index 8625675..d4c7ebd 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_lastprivate_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_lastprivate_codegen.cpp
@@ -834,9 +834,26 @@ int main() {
 // CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK5-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, i32 5, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK5-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK5-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 5, i32* [[TMP35]], align 4
+// CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP32]], i8*** [[TMP36]], align 8
+// CHECK5-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP33]], i8*** [[TMP37]], align 8
+// CHECK5-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 8
+// CHECK5-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 8
+// CHECK5-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP40]], align 8
+// CHECK5-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP41]], align 8
+// CHECK5-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK5-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106([2 x i32]* [[VEC]], i64 [[TMP3]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -844,18 +861,18 @@ int main() {
 // CHECK5-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK5-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK5-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK5-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK5-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK5:       arraydestroy.body:
-// CHECK5-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP36]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK5-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP44]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK5-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK5-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK5-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK5-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK5:       arraydestroy.done3:
 // CHECK5-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK5-NEXT:    [[TMP37:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP37]]
+// CHECK5-NEXT:    [[TMP45:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP45]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1313,27 +1330,44 @@ int main() {
 // CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK5-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK5-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 8
+// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP32]], align 8
+// CHECK5-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK5-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP34]], align 8
+// CHECK5-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK5-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50([2 x i32]* [[VEC]], i64 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
 // CHECK5-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK5-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK5-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK5-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK5:       arraydestroy.body:
-// CHECK5-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK5-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK5-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK5-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK5-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK5-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK5:       arraydestroy.done2:
 // CHECK5-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK5-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP30]]
+// CHECK5-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1857,9 +1891,26 @@ int main() {
 // CHECK7-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK7-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, i32 5, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK7-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK7-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 5, i32* [[TMP35]], align 4
+// CHECK7-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP32]], i8*** [[TMP36]], align 4
+// CHECK7-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP33]], i8*** [[TMP37]], align 4
+// CHECK7-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 4
+// CHECK7-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 4
+// CHECK7-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP40]], align 4
+// CHECK7-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP41]], align 4
+// CHECK7-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK7-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106([2 x i32]* [[VEC]], i32 [[TMP3]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1867,18 +1918,18 @@ int main() {
 // CHECK7-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK7-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK7-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK7-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK7-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK7:       arraydestroy.body:
-// CHECK7-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP36]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK7-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP44]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK7-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK7-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK7-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK7-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK7:       arraydestroy.done2:
 // CHECK7-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK7-NEXT:    [[TMP37:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK7-NEXT:    ret i32 [[TMP37]]
+// CHECK7-NEXT:    [[TMP45:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK7-NEXT:    ret i32 [[TMP45]]
 //
 //
 // CHECK7-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -2319,27 +2370,44 @@ int main() {
 // CHECK7-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK7-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK7-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK7-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK7-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 4
+// CHECK7-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK7-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP32]], align 4
+// CHECK7-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK7-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP34]], align 4
+// CHECK7-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK7-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50([2 x i32]* [[VEC]], i32 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK7:       omp_offload.cont:
 // CHECK7-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK7-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK7-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK7-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK7:       arraydestroy.body:
-// CHECK7-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK7-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK7-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK7-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK7-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK7-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK7:       arraydestroy.done2:
 // CHECK7-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK7-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK7-NEXT:    ret i32 [[TMP30]]
+// CHECK7-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK7-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK7-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_order_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_order_codegen.cpp
index 3e2e130..446623d 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_order_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_order_codegen.cpp
@@ -26,9 +26,26 @@ void gtid_test() {
 // CHECK1-NEXT:  entry:
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l16.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l16.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l16() #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_private_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_private_codegen.cpp
index f1e426b..09587e5 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_private_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_private_codegen.cpp
@@ -300,9 +300,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l124.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l124.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l124() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -547,27 +564,44 @@ int main() {
 // CHECK1-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 8
 // CHECK1-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 8
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l80.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK1-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l80() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP4]]
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev
@@ -993,9 +1027,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l124.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l124.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l124() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1234,27 +1285,44 @@ int main() {
 // CHECK3-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 4
 // CHECK3-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l80.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK3-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l80() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP4]]
+// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_proc_bind_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_proc_bind_codegen.cpp
index 509ab7d..b398c4c 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_proc_bind_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_proc_bind_codegen.cpp
@@ -56,21 +56,55 @@ int main() {
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l36.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l36() #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l38.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l38.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l38() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK1-NEXT:    ret i32 [[CALL]]
 //
@@ -356,9 +390,26 @@ int main() {
 // CHECK1-NEXT:  entry:
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l30.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l30.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l30() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_reduction_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_reduction_codegen.cpp
index 6528be7..036a7cb 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_reduction_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_reduction_codegen.cpp
@@ -110,9 +110,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l66.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l66.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l66(i32* @_ZZ4mainE5sivar) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -390,9 +407,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
-// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
+// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i32* [[T_VAR]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -673,9 +707,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l66.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l66.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l66(i32* @_ZZ4mainE5sivar) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -949,9 +1000,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
-// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
+// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i32* [[T_VAR]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_schedule_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_schedule_codegen.cpp
index 2795e43..f17ce11 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_schedule_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_schedule_codegen.cpp
@@ -214,18 +214,18 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -240,93 +240,178 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK1:       omp_offload.failed16:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK1:       omp_offload.cont15:
-// CHECK1-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 8
-// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 8
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP31]], align 8
-// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK1:       omp_offload.cont17:
+// CHECK1-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 8
+// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 8
+// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP55]], align 8
+// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK1-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK1:       omp_offload.failed21:
+// CHECK1-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK1-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK1-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 8
+// CHECK1-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 8
+// CHECK1-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 8
+// CHECK1-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 8
+// CHECK1-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP64]], align 8
+// CHECK1-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP65]], align 8
+// CHECK1-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK1-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK1-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK1:       omp_offload.failed24:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK1:       omp_offload.cont22:
-// CHECK1-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 8
-// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 8
-// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK1:       omp_offload.cont25:
+// CHECK1-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 8
+// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP72]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK1-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK1:       omp_offload.failed28:
+// CHECK1-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK1-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 8
+// CHECK1-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 8
+// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 8
+// CHECK1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 8
+// CHECK1-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP81]], align 8
+// CHECK1-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP82]], align 8
+// CHECK1-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK1-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK1-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK1:       omp_offload.failed32:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK1:       omp_offload.cont29:
-// CHECK1-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP45]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK1:       omp_offload.cont33:
+// CHECK1-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i64 0, i64 0
+// CHECK1-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36
@@ -1128,18 +1213,18 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -1154,93 +1239,178 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK3:       omp_offload.failed8:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK3:       omp_offload.cont8:
-// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK3:       omp_offload.cont9:
+// CHECK3-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK3-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK3:       omp_offload.failed14:
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK3:       omp_offload.cont15:
-// CHECK3-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 4
-// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 4
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP31]], align 4
-// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 4
+// CHECK3-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 4
+// CHECK3-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP55]], align 4
+// CHECK3-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK3-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK3:       omp_offload.failed21:
+// CHECK3-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK3-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK3-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 4
+// CHECK3-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 4
+// CHECK3-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 4
+// CHECK3-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 4
+// CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP64]], align 4
+// CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP65]], align 4
+// CHECK3-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK3-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK3-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK3:       omp_offload.failed24:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK3:       omp_offload.cont22:
-// CHECK3-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 4
-// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 4
-// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP40]], align 4
-// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK3:       omp_offload.cont25:
+// CHECK3-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 4
+// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP72]], align 4
+// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK3-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK3:       omp_offload.failed28:
+// CHECK3-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK3-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK3-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 4
+// CHECK3-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 4
+// CHECK3-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 4
+// CHECK3-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 4
+// CHECK3-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP81]], align 4
+// CHECK3-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP82]], align 4
+// CHECK3-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK3-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK3-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK3:       omp_offload.failed32:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK3:       omp_offload.cont29:
-// CHECK3-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP45]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK3:       omp_offload.cont33:
+// CHECK3-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36
@@ -2015,18 +2185,18 @@ int main (int argc, char **argv) {
 // CHECK5-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK5-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK5-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK5-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -2041,93 +2211,178 @@ int main (int argc, char **argv) {
 // CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK5-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK5-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK5-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK5-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK5-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
 // CHECK5-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK5-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK5-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK5-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK5-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK5:       omp_offload.failed7:
+// CHECK5-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK5-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK5-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK5-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK5:       omp_offload.failed8:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK5:       omp_offload.cont8:
-// CHECK5-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK5-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK5-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK5-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK5:       omp_offload.cont9:
+// CHECK5-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK5-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK5-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK5-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK5:       omp_offload.failed14:
+// CHECK5-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK5-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK5-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK5-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK5-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK5-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK5-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK5-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK5-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK5-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK5-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK5:       omp_offload.failed16:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK5:       omp_offload.cont15:
-// CHECK5-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 8
-// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 8
-// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP31]], align 8
-// CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK5:       omp_offload.cont17:
+// CHECK5-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 8
+// CHECK5-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 8
+// CHECK5-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP55]], align 8
+// CHECK5-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK5-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK5:       omp_offload.failed21:
+// CHECK5-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK5-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK5-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 8
+// CHECK5-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 8
+// CHECK5-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 8
+// CHECK5-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 8
+// CHECK5-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP64]], align 8
+// CHECK5-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP65]], align 8
+// CHECK5-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK5-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK5-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK5:       omp_offload.failed24:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK5:       omp_offload.cont22:
-// CHECK5-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 8
-// CHECK5-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 8
-// CHECK5-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK5-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK5:       omp_offload.cont25:
+// CHECK5-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 8
+// CHECK5-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 8
+// CHECK5-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP72]], align 8
+// CHECK5-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK5-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK5:       omp_offload.failed28:
+// CHECK5-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK5-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK5-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 8
+// CHECK5-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 8
+// CHECK5-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 8
+// CHECK5-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 8
+// CHECK5-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP81]], align 8
+// CHECK5-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP82]], align 8
+// CHECK5-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK5-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK5-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK5:       omp_offload.failed32:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK5:       omp_offload.cont29:
-// CHECK5-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i64 0, i64 0
-// CHECK5-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP45]]
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK5:       omp_offload.cont33:
+// CHECK5-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i64 0, i64 0
+// CHECK5-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36
@@ -2929,18 +3184,18 @@ int main (int argc, char **argv) {
 // CHECK7-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK7-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK7-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK7-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK7-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -2955,93 +3210,178 @@ int main (int argc, char **argv) {
 // CHECK7-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK7-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK7-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK7-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK7-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK7-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK7-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK7-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK7-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK7-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK7-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK7:       omp_offload.cont:
 // CHECK7-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK7-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK7-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK7-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK7-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK7-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK7-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK7-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK7-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK7-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK7:       omp_offload.failed7:
+// CHECK7-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK7-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK7-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK7-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK7-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK7-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK7-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK7-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK7-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK7:       omp_offload.failed8:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK7:       omp_offload.cont8:
-// CHECK7-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK7-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK7-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK7-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK7-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK7:       omp_offload.cont9:
+// CHECK7-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK7-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK7-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK7-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK7-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK7-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK7:       omp_offload.failed14:
+// CHECK7-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK7-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK7-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK7-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK7-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK7-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK7-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK7-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK7-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK7-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK7-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK7:       omp_offload.failed16:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK7:       omp_offload.cont15:
-// CHECK7-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 4
-// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK7-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 4
-// CHECK7-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP31]], align 4
-// CHECK7-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK7:       omp_offload.cont17:
+// CHECK7-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 4
+// CHECK7-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK7-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 4
+// CHECK7-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP55]], align 4
+// CHECK7-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK7-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK7:       omp_offload.failed21:
+// CHECK7-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK7-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK7-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 4
+// CHECK7-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 4
+// CHECK7-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 4
+// CHECK7-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 4
+// CHECK7-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP64]], align 4
+// CHECK7-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP65]], align 4
+// CHECK7-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK7-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK7-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK7:       omp_offload.failed24:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK7:       omp_offload.cont22:
-// CHECK7-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 4
-// CHECK7-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK7-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 4
-// CHECK7-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP40]], align 4
-// CHECK7-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK7:       omp_offload.cont25:
+// CHECK7-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 4
+// CHECK7-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK7-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 4
+// CHECK7-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP72]], align 4
+// CHECK7-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK7-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK7:       omp_offload.failed28:
+// CHECK7-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK7-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK7-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 4
+// CHECK7-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 4
+// CHECK7-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 4
+// CHECK7-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 4
+// CHECK7-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP81]], align 4
+// CHECK7-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP82]], align 4
+// CHECK7-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK7-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK7-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK7:       omp_offload.failed32:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK7:       omp_offload.cont29:
-// CHECK7-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK7-NEXT:    ret i32 [[TMP45]]
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK7:       omp_offload.cont33:
+// CHECK7-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK7-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK7-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36
@@ -3822,34 +4162,34 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[_TMP9:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
-// CHECK13-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES26:%.*]] = alloca [4 x i64], align 8
+// CHECK13-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[N_CASTED35:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS37:%.*]] = alloca [3 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS38:%.*]] = alloca [3 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS39:%.*]] = alloca [3 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES40:%.*]] = alloca [3 x i64], align 8
-// CHECK13-NEXT:    [[_TMP41:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_42:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_43:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_50:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[N_CASTED51:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR__CASTED53:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS55:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS56:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS57:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES58:%.*]] = alloca [4 x i64], align 8
-// CHECK13-NEXT:    [[_TMP59:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_60:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_61:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_29:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[N_CASTED37:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS39:%.*]] = alloca [3 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS40:%.*]] = alloca [3 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS41:%.*]] = alloca [3 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES42:%.*]] = alloca [3 x i64], align 8
+// CHECK13-NEXT:    [[_TMP43:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_44:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_45:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_53:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[N_CASTED54:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR__CASTED56:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS58:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS59:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS60:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES61:%.*]] = alloca [4 x i64], align 8
+// CHECK13-NEXT:    [[_TMP62:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_63:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_64:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK13-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK13-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -3908,264 +4248,349 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK13-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK13-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK13-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK13-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK13-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK13-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK13-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK13-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK13-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK13-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK13-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK13-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       omp_offload.cont:
-// CHECK13-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK13-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK13-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK13-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
-// CHECK13-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK13-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK13-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK13-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK13-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK13-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK13-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK13-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK13-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK13-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK13-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK13-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK13-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK13-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK13-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK13-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
+// CHECK13-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK13-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK13-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK13-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK13-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK13-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK13-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK13-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK13-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK13-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK13-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK13-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK13-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK13-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK13-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK13-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK13-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK13-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP58]])
-// CHECK13-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK13-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK13:       omp_offload.failed16:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK13:       omp_offload.cont17:
-// CHECK13-NEXT:    [[TMP61:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    store i32 [[TMP61]], i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK13-NEXT:    [[TMP62:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP62]], i32* [[CONV20]], align 4
-// CHECK13-NEXT:    [[TMP63:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK13-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK13-NEXT:    [[CONV21:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP64]], i32* [[CONV21]], align 4
-// CHECK13-NEXT:    [[TMP65:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK13-NEXT:    [[TMP66:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP67:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP67]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
-// CHECK13-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP63]], i64* [[TMP69]], align 8
-// CHECK13-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP63]], i64* [[TMP71]], align 8
-// CHECK13-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP72]], align 8
-// CHECK13-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP74]], align 8
-// CHECK13-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP76]], align 8
-// CHECK13-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP77]], align 8
-// CHECK13-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP79]], align 8
-// CHECK13-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 8
-// CHECK13-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 2
-// CHECK13-NEXT:    store i64 [[TMP66]], i64* [[TMP82]], align 8
-// CHECK13-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP83]], align 8
-// CHECK13-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK13-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK13-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP66]])
+// CHECK13-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK13-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK13-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK13-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK13-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK13-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK13-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK13-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK13-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK13-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK13-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK13:       omp_offload.failed17:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK13:       omp_offload.cont18:
+// CHECK13-NEXT:    [[TMP77:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    store i32 [[TMP77]], i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK13-NEXT:    [[TMP78:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP78]], i32* [[CONV21]], align 4
+// CHECK13-NEXT:    [[TMP79:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK13-NEXT:    [[TMP80:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK13-NEXT:    [[CONV22:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP80]], i32* [[CONV22]], align 4
+// CHECK13-NEXT:    [[TMP81:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK13-NEXT:    [[TMP82:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP83:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES26]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP83]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
+// CHECK13-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
 // CHECK13-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP65]], i64* [[TMP85]], align 8
-// CHECK13-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK13-NEXT:    store i64 [[TMP79]], i64* [[TMP85]], align 8
+// CHECK13-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
 // CHECK13-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP65]], i64* [[TMP87]], align 8
-// CHECK13-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK13-NEXT:    store i64 [[TMP79]], i64* [[TMP87]], align 8
+// CHECK13-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 0
 // CHECK13-NEXT:    store i8* null, i8** [[TMP88]], align 8
-// CHECK13-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP92:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP92]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK13-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK13-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP93]], 0
-// CHECK13-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
-// CHECK13-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
-// CHECK13-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK13-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK13-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP94]], 1
-// CHECK13-NEXT:    [[TMP95:%.*]] = zext i32 [[ADD32]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP95]])
-// CHECK13-NEXT:    [[TMP96:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, i32 4, i8** [[TMP89]], i8** [[TMP90]], i64* [[TMP91]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP97:%.*]] = icmp ne i32 [[TMP96]], 0
-// CHECK13-NEXT:    br i1 [[TMP97]], label [[OMP_OFFLOAD_FAILED33:%.*]], label [[OMP_OFFLOAD_CONT34:%.*]]
-// CHECK13:       omp_offload.failed33:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i64 [[TMP63]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP65]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT34]]
-// CHECK13:       omp_offload.cont34:
-// CHECK13-NEXT:    [[TMP98:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    [[CONV36:%.*]] = bitcast i64* [[N_CASTED35]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP98]], i32* [[CONV36]], align 4
-// CHECK13-NEXT:    [[TMP99:%.*]] = load i64, i64* [[N_CASTED35]], align 8
-// CHECK13-NEXT:    [[TMP100:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP101:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES40]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP101]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
-// CHECK13-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP90]], align 8
+// CHECK13-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP92]], align 8
+// CHECK13-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP93]], align 8
+// CHECK13-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP95]], align 8
+// CHECK13-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 8
+// CHECK13-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 2
+// CHECK13-NEXT:    store i64 [[TMP82]], i64* [[TMP98]], align 8
+// CHECK13-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP99]], align 8
+// CHECK13-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP81]], i64* [[TMP101]], align 8
+// CHECK13-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 3
 // CHECK13-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP99]], i64* [[TMP103]], align 8
-// CHECK13-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP99]], i64* [[TMP105]], align 8
-// CHECK13-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS39]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP106]], align 8
-// CHECK13-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP108]], align 8
-// CHECK13-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP110]], align 8
-// CHECK13-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS39]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP111]], align 8
-// CHECK13-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP113]], align 8
-// CHECK13-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP115]], align 8
-// CHECK13-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES40]], i32 0, i32 2
-// CHECK13-NEXT:    store i64 [[TMP100]], i64* [[TMP116]], align 8
-// CHECK13-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS39]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP117]], align 8
-// CHECK13-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES40]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP121:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP121]], i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK13-NEXT:    [[TMP122:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK13-NEXT:    [[SUB44:%.*]] = sub nsw i32 [[TMP122]], 0
-// CHECK13-NEXT:    [[DIV45:%.*]] = sdiv i32 [[SUB44]], 1
-// CHECK13-NEXT:    [[SUB46:%.*]] = sub nsw i32 [[DIV45]], 1
-// CHECK13-NEXT:    store i32 [[SUB46]], i32* [[DOTCAPTURE_EXPR_43]], align 4
-// CHECK13-NEXT:    [[TMP123:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_43]], align 4
-// CHECK13-NEXT:    [[ADD47:%.*]] = add nsw i32 [[TMP123]], 1
-// CHECK13-NEXT:    [[TMP124:%.*]] = zext i32 [[ADD47]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP124]])
-// CHECK13-NEXT:    [[TMP125:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, i32 3, i8** [[TMP118]], i8** [[TMP119]], i64* [[TMP120]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP126:%.*]] = icmp ne i32 [[TMP125]], 0
-// CHECK13-NEXT:    br i1 [[TMP126]], label [[OMP_OFFLOAD_FAILED48:%.*]], label [[OMP_OFFLOAD_CONT49:%.*]]
-// CHECK13:       omp_offload.failed48:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i64 [[TMP99]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT49]]
-// CHECK13:       omp_offload.cont49:
-// CHECK13-NEXT:    [[TMP127:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    store i32 [[TMP127]], i32* [[DOTCAPTURE_EXPR_50]], align 4
-// CHECK13-NEXT:    [[TMP128:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    [[CONV52:%.*]] = bitcast i64* [[N_CASTED51]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP128]], i32* [[CONV52]], align 4
-// CHECK13-NEXT:    [[TMP129:%.*]] = load i64, i64* [[N_CASTED51]], align 8
-// CHECK13-NEXT:    [[TMP130:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_50]], align 4
-// CHECK13-NEXT:    [[CONV54:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED53]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP130]], i32* [[CONV54]], align 4
-// CHECK13-NEXT:    [[TMP131:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED53]], align 8
-// CHECK13-NEXT:    [[TMP132:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP133:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES58]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP133]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
-// CHECK13-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP129]], i64* [[TMP135]], align 8
-// CHECK13-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP129]], i64* [[TMP137]], align 8
-// CHECK13-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP138]], align 8
-// CHECK13-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP140]], align 8
-// CHECK13-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP142]], align 8
-// CHECK13-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP143]], align 8
-// CHECK13-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP145:%.*]] = bitcast i8** [[TMP144]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP145]], align 8
-// CHECK13-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP147]], align 8
-// CHECK13-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES58]], i32 0, i32 2
-// CHECK13-NEXT:    store i64 [[TMP132]], i64* [[TMP148]], align 8
-// CHECK13-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP149]], align 8
-// CHECK13-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 3
-// CHECK13-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP131]], i64* [[TMP151]], align 8
-// CHECK13-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 3
-// CHECK13-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP131]], i64* [[TMP153]], align 8
-// CHECK13-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 3
-// CHECK13-NEXT:    store i8* null, i8** [[TMP154]], align 8
-// CHECK13-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES58]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP158:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP158]], i32* [[DOTCAPTURE_EXPR_60]], align 4
-// CHECK13-NEXT:    [[TMP159:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_60]], align 4
-// CHECK13-NEXT:    [[SUB62:%.*]] = sub nsw i32 [[TMP159]], 0
-// CHECK13-NEXT:    [[DIV63:%.*]] = sdiv i32 [[SUB62]], 1
-// CHECK13-NEXT:    [[SUB64:%.*]] = sub nsw i32 [[DIV63]], 1
-// CHECK13-NEXT:    store i32 [[SUB64]], i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK13-NEXT:    [[TMP160:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK13-NEXT:    [[ADD65:%.*]] = add nsw i32 [[TMP160]], 1
-// CHECK13-NEXT:    [[TMP161:%.*]] = zext i32 [[ADD65]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP161]])
-// CHECK13-NEXT:    [[TMP162:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, i32 4, i8** [[TMP155]], i8** [[TMP156]], i64* [[TMP157]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP163:%.*]] = icmp ne i32 [[TMP162]], 0
-// CHECK13-NEXT:    br i1 [[TMP163]], label [[OMP_OFFLOAD_FAILED66:%.*]], label [[OMP_OFFLOAD_CONT67:%.*]]
-// CHECK13:       omp_offload.failed66:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i64 [[TMP129]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP131]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT67]]
-// CHECK13:       omp_offload.cont67:
-// CHECK13-NEXT:    [[TMP164:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK13-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP164]])
+// CHECK13-NEXT:    store i64 [[TMP81]], i64* [[TMP103]], align 8
+// CHECK13-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 3
+// CHECK13-NEXT:    store i8* null, i8** [[TMP104]], align 8
+// CHECK13-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP108:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP108]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK13-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK13-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[TMP109]], 0
+// CHECK13-NEXT:    [[DIV31:%.*]] = sdiv i32 [[SUB30]], 1
+// CHECK13-NEXT:    [[SUB32:%.*]] = sub nsw i32 [[DIV31]], 1
+// CHECK13-NEXT:    store i32 [[SUB32]], i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK13-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK13-NEXT:    [[ADD33:%.*]] = add nsw i32 [[TMP110]], 1
+// CHECK13-NEXT:    [[TMP111:%.*]] = zext i32 [[ADD33]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP111]])
+// CHECK13-NEXT:    [[KERNEL_ARGS34:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP112]], align 4
+// CHECK13-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 4, i32* [[TMP113]], align 4
+// CHECK13-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK13-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 8
+// CHECK13-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP107]], i64** [[TMP116]], align 8
+// CHECK13-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP117]], align 8
+// CHECK13-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK13-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP119]], align 8
+// CHECK13-NEXT:    [[TMP120:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]])
+// CHECK13-NEXT:    [[TMP121:%.*]] = icmp ne i32 [[TMP120]], 0
+// CHECK13-NEXT:    br i1 [[TMP121]], label [[OMP_OFFLOAD_FAILED35:%.*]], label [[OMP_OFFLOAD_CONT36:%.*]]
+// CHECK13:       omp_offload.failed35:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i64 [[TMP79]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP81]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT36]]
+// CHECK13:       omp_offload.cont36:
+// CHECK13-NEXT:    [[TMP122:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[CONV38:%.*]] = bitcast i64* [[N_CASTED37]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP122]], i32* [[CONV38]], align 4
+// CHECK13-NEXT:    [[TMP123:%.*]] = load i64, i64* [[N_CASTED37]], align 8
+// CHECK13-NEXT:    [[TMP124:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP125:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES42]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP125]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
+// CHECK13-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP123]], i64* [[TMP127]], align 8
+// CHECK13-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP123]], i64* [[TMP129]], align 8
+// CHECK13-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS41]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP130]], align 8
+// CHECK13-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP132]], align 8
+// CHECK13-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP134]], align 8
+// CHECK13-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS41]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP135]], align 8
+// CHECK13-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP137]], align 8
+// CHECK13-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP139]], align 8
+// CHECK13-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES42]], i32 0, i32 2
+// CHECK13-NEXT:    store i64 [[TMP124]], i64* [[TMP140]], align 8
+// CHECK13-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS41]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP141]], align 8
+// CHECK13-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES42]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP145:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP145]], i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK13-NEXT:    [[TMP146:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK13-NEXT:    [[SUB46:%.*]] = sub nsw i32 [[TMP146]], 0
+// CHECK13-NEXT:    [[DIV47:%.*]] = sdiv i32 [[SUB46]], 1
+// CHECK13-NEXT:    [[SUB48:%.*]] = sub nsw i32 [[DIV47]], 1
+// CHECK13-NEXT:    store i32 [[SUB48]], i32* [[DOTCAPTURE_EXPR_45]], align 4
+// CHECK13-NEXT:    [[TMP147:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_45]], align 4
+// CHECK13-NEXT:    [[ADD49:%.*]] = add nsw i32 [[TMP147]], 1
+// CHECK13-NEXT:    [[TMP148:%.*]] = zext i32 [[ADD49]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP148]])
+// CHECK13-NEXT:    [[KERNEL_ARGS50:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP149]], align 4
+// CHECK13-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 3, i32* [[TMP150]], align 4
+// CHECK13-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP142]], i8*** [[TMP151]], align 8
+// CHECK13-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP143]], i8*** [[TMP152]], align 8
+// CHECK13-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP144]], i64** [[TMP153]], align 8
+// CHECK13-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP154]], align 8
+// CHECK13-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP155]], align 8
+// CHECK13-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP156]], align 8
+// CHECK13-NEXT:    [[TMP157:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]])
+// CHECK13-NEXT:    [[TMP158:%.*]] = icmp ne i32 [[TMP157]], 0
+// CHECK13-NEXT:    br i1 [[TMP158]], label [[OMP_OFFLOAD_FAILED51:%.*]], label [[OMP_OFFLOAD_CONT52:%.*]]
+// CHECK13:       omp_offload.failed51:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i64 [[TMP123]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT52]]
+// CHECK13:       omp_offload.cont52:
+// CHECK13-NEXT:    [[TMP159:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    store i32 [[TMP159]], i32* [[DOTCAPTURE_EXPR_53]], align 4
+// CHECK13-NEXT:    [[TMP160:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[CONV55:%.*]] = bitcast i64* [[N_CASTED54]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP160]], i32* [[CONV55]], align 4
+// CHECK13-NEXT:    [[TMP161:%.*]] = load i64, i64* [[N_CASTED54]], align 8
+// CHECK13-NEXT:    [[TMP162:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_53]], align 4
+// CHECK13-NEXT:    [[CONV57:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED56]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP162]], i32* [[CONV57]], align 4
+// CHECK13-NEXT:    [[TMP163:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED56]], align 8
+// CHECK13-NEXT:    [[TMP164:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP165:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES61]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP165]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
+// CHECK13-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP161]], i64* [[TMP167]], align 8
+// CHECK13-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP161]], i64* [[TMP169]], align 8
+// CHECK13-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP170]], align 8
+// CHECK13-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP172:%.*]] = bitcast i8** [[TMP171]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP172]], align 8
+// CHECK13-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP174:%.*]] = bitcast i8** [[TMP173]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP174]], align 8
+// CHECK13-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP175]], align 8
+// CHECK13-NEXT:    [[TMP176:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP177:%.*]] = bitcast i8** [[TMP176]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP177]], align 8
+// CHECK13-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP179]], align 8
+// CHECK13-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES61]], i32 0, i32 2
+// CHECK13-NEXT:    store i64 [[TMP164]], i64* [[TMP180]], align 8
+// CHECK13-NEXT:    [[TMP181:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP181]], align 8
+// CHECK13-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP183:%.*]] = bitcast i8** [[TMP182]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP163]], i64* [[TMP183]], align 8
+// CHECK13-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP185:%.*]] = bitcast i8** [[TMP184]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP163]], i64* [[TMP185]], align 8
+// CHECK13-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 3
+// CHECK13-NEXT:    store i8* null, i8** [[TMP186]], align 8
+// CHECK13-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES61]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP190:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP190]], i32* [[DOTCAPTURE_EXPR_63]], align 4
+// CHECK13-NEXT:    [[TMP191:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_63]], align 4
+// CHECK13-NEXT:    [[SUB65:%.*]] = sub nsw i32 [[TMP191]], 0
+// CHECK13-NEXT:    [[DIV66:%.*]] = sdiv i32 [[SUB65]], 1
+// CHECK13-NEXT:    [[SUB67:%.*]] = sub nsw i32 [[DIV66]], 1
+// CHECK13-NEXT:    store i32 [[SUB67]], i32* [[DOTCAPTURE_EXPR_64]], align 4
+// CHECK13-NEXT:    [[TMP192:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_64]], align 4
+// CHECK13-NEXT:    [[ADD68:%.*]] = add nsw i32 [[TMP192]], 1
+// CHECK13-NEXT:    [[TMP193:%.*]] = zext i32 [[ADD68]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP193]])
+// CHECK13-NEXT:    [[KERNEL_ARGS69:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP194]], align 4
+// CHECK13-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 4, i32* [[TMP195]], align 4
+// CHECK13-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP187]], i8*** [[TMP196]], align 8
+// CHECK13-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP188]], i8*** [[TMP197]], align 8
+// CHECK13-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP189]], i64** [[TMP198]], align 8
+// CHECK13-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP199]], align 8
+// CHECK13-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP200]], align 8
+// CHECK13-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP201]], align 8
+// CHECK13-NEXT:    [[TMP202:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]])
+// CHECK13-NEXT:    [[TMP203:%.*]] = icmp ne i32 [[TMP202]], 0
+// CHECK13-NEXT:    br i1 [[TMP203]], label [[OMP_OFFLOAD_FAILED70:%.*]], label [[OMP_OFFLOAD_CONT71:%.*]]
+// CHECK13:       omp_offload.failed70:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i64 [[TMP161]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP163]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT71]]
+// CHECK13:       omp_offload.cont71:
+// CHECK13-NEXT:    [[TMP204:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK13-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP204]])
 // CHECK13-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK13-NEXT:    [[TMP165:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK13-NEXT:    call void @llvm.stackrestore(i8* [[TMP165]])
-// CHECK13-NEXT:    [[TMP166:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK13-NEXT:    ret i32 [[TMP166]]
+// CHECK13-NEXT:    [[TMP205:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK13-NEXT:    call void @llvm.stackrestore(i8* [[TMP205]])
+// CHECK13-NEXT:    [[TMP206:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK13-NEXT:    ret i32 [[TMP206]]
 //
 //
 // CHECK13-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139
@@ -5350,20 +5775,20 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 8
-// CHECK13-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR__CASTED20:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 8
+// CHECK13-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR__CASTED23:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS25:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS26:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS27:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[_TMP28:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK13-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK13-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -5377,113 +5802,198 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK13-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK13-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK13-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK13-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK13-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK13-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK13-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK13-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK13-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       omp_offload.cont:
-// CHECK13-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK13-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK13-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK13-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK13-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK13-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK13:       omp_offload.failed5:
+// CHECK13-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK13-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK13-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK13-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK13-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK13-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK13-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK13-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK13-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK13-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK13-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK13:       omp_offload.failed6:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK13:       omp_offload.cont6:
-// CHECK13-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    store i32 [[TMP18]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK13-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK13:       omp_offload.cont7:
+// CHECK13-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    store i32 [[TMP34]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK13-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
 // CHECK13-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP19]], i32* [[CONV]], align 4
-// CHECK13-NEXT:    [[TMP20:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK13-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP22]], align 8
-// CHECK13-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP24]], align 8
-// CHECK13-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP25]], align 8
-// CHECK13-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP20]], i64* [[TMP27]], align 8
-// CHECK13-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP20]], i64* [[TMP29]], align 8
-// CHECK13-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP30]], align 8
-// CHECK13-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, i32 2, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK13-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK13:       omp_offload.failed11:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i64 [[TMP20]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK13:       omp_offload.cont12:
-// CHECK13-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP36]], align 8
-// CHECK13-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 [[TMP35]], i32* [[CONV]], align 4
+// CHECK13-NEXT:    [[TMP36:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK13-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
 // CHECK13-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [10 x i32]**
 // CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP38]], align 8
-// CHECK13-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP39]], align 8
-// CHECK13-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP40]], align 8
+// CHECK13-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP41]], align 8
+// CHECK13-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP36]], i64* [[TMP43]], align 8
+// CHECK13-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP36]], i64* [[TMP45]], align 8
+// CHECK13-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP46]], align 8
+// CHECK13-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, i32 1, i8** [[TMP40]], i8** [[TMP41]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK13-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK13:       omp_offload.failed17:
+// CHECK13-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP49]], align 4
+// CHECK13-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 2, i32* [[TMP50]], align 4
+// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 8
+// CHECK13-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP48]], i8*** [[TMP52]], align 8
+// CHECK13-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP54]], align 8
+// CHECK13-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK13-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP56]], align 8
+// CHECK13-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK13-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
+// CHECK13-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK13:       omp_offload.failed13:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i64 [[TMP36]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK13:       omp_offload.cont14:
+// CHECK13-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP60]], align 8
+// CHECK13-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP62]], align 8
+// CHECK13-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP63]], align 8
+// CHECK13-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
+// CHECK13-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK13-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK13-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 8
+// CHECK13-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 8
+// CHECK13-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP70]], align 8
+// CHECK13-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP71]], align 8
+// CHECK13-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP72]], align 8
+// CHECK13-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK13-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK13-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK13-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK13:       omp_offload.failed20:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK13:       omp_offload.cont18:
-// CHECK13-NEXT:    [[TMP44:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    store i32 [[TMP44]], i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK13-NEXT:    [[TMP45:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK13-NEXT:    [[CONV21:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED20]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP45]], i32* [[CONV21]], align 4
-// CHECK13-NEXT:    [[TMP46:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED20]], align 8
-// CHECK13-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP48]], align 8
-// CHECK13-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP50]], align 8
-// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK13-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP46]], i64* [[TMP53]], align 8
-// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP46]], i64* [[TMP55]], align 8
-// CHECK13-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP56]], align 8
-// CHECK13-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK13:       omp_offload.cont21:
+// CHECK13-NEXT:    [[TMP76:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    store i32 [[TMP76]], i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK13-NEXT:    [[TMP77:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK13-NEXT:    [[CONV24:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED23]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP77]], i32* [[CONV24]], align 4
+// CHECK13-NEXT:    [[TMP78:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED23]], align 8
+// CHECK13-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP80]], align 8
+// CHECK13-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP82]], align 8
+// CHECK13-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP83]], align 8
+// CHECK13-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP78]], i64* [[TMP85]], align 8
+// CHECK13-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP78]], i64* [[TMP87]], align 8
+// CHECK13-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP88]], align 8
+// CHECK13-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, i32 2, i8** [[TMP57]], i8** [[TMP58]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK13-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED26:%.*]], label [[OMP_OFFLOAD_CONT27:%.*]]
-// CHECK13:       omp_offload.failed26:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i64 [[TMP46]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT27]]
-// CHECK13:       omp_offload.cont27:
+// CHECK13-NEXT:    [[KERNEL_ARGS29:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP91]], align 4
+// CHECK13-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 2, i32* [[TMP92]], align 4
+// CHECK13-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP89]], i8*** [[TMP93]], align 8
+// CHECK13-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP90]], i8*** [[TMP94]], align 8
+// CHECK13-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP95]], align 8
+// CHECK13-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP96]], align 8
+// CHECK13-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP97]], align 8
+// CHECK13-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP98]], align 8
+// CHECK13-NEXT:    [[TMP99:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]])
+// CHECK13-NEXT:    [[TMP100:%.*]] = icmp ne i32 [[TMP99]], 0
+// CHECK13-NEXT:    br i1 [[TMP100]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
+// CHECK13:       omp_offload.failed30:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i64 [[TMP78]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
+// CHECK13:       omp_offload.cont31:
 // CHECK13-NEXT:    ret i32 0
 //
 //
@@ -6327,34 +6837,34 @@ int main (int argc, char **argv) {
 // CHECK15-NEXT:    [[_TMP8:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_17:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[N_CASTED19:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
-// CHECK15-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES23:%.*]] = alloca [4 x i64], align 4
+// CHECK15-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[N_CASTED32:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [3 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [3 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [3 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES36:%.*]] = alloca [3 x i64], align 4
-// CHECK15-NEXT:    [[_TMP37:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_38:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_39:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_46:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[N_CASTED47:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR__CASTED48:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS49:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS50:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS51:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES52:%.*]] = alloca [4 x i64], align 4
-// CHECK15-NEXT:    [[_TMP53:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_54:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_55:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[N_CASTED34:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS35:%.*]] = alloca [3 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS36:%.*]] = alloca [3 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS37:%.*]] = alloca [3 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES38:%.*]] = alloca [3 x i64], align 4
+// CHECK15-NEXT:    [[_TMP39:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_40:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_41:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_49:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[N_CASTED50:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR__CASTED51:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS52:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS53:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS54:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES55:%.*]] = alloca [4 x i64], align 4
+// CHECK15-NEXT:    [[_TMP56:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_57:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_58:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK15-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK15-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -6412,262 +6922,347 @@ int main (int argc, char **argv) {
 // CHECK15-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK15-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK15-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK15-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK15-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK15-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK15-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK15-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK15-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK15-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK15-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK15-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK15-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK15-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK15-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK15-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK15:       omp_offload.failed:
 // CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK15:       omp_offload.cont:
-// CHECK15-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK15-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK15-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK15-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK15-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
-// CHECK15-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK15-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK15-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK15-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK15-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK15-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK15-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK15-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK15-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK15-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK15-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK15-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK15-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK15-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK15-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK15-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK15-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK15-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK15-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK15-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
+// CHECK15-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK15-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK15-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK15-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK15-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK15-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK15-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK15-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK15-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK15-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK15-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK15-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK15-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK15-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK15-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK15-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK15-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK15-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK15-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK15-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK15-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP59]])
-// CHECK15-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK15-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK15:       omp_offload.failed15:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK15:       omp_offload.cont16:
-// CHECK15-NEXT:    [[TMP62:%.*]] = load i32, i32* [[M]], align 4
-// CHECK15-NEXT:    store i32 [[TMP62]], i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK15-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP63]], i32* [[N_CASTED18]], align 4
-// CHECK15-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N_CASTED18]], align 4
-// CHECK15-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK15-NEXT:    store i32 [[TMP65]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK15-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK15-NEXT:    [[TMP67:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK15-NEXT:    [[TMP68:%.*]] = sext i32 [[TMP67]] to i64
-// CHECK15-NEXT:    [[TMP69:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
-// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP69]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
-// CHECK15-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP64]], i32* [[TMP71]], align 4
-// CHECK15-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP64]], i32* [[TMP73]], align 4
-// CHECK15-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP74]], align 4
-// CHECK15-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP76]], align 4
-// CHECK15-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP78]], align 4
-// CHECK15-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP79]], align 4
-// CHECK15-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 4
-// CHECK15-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 4
-// CHECK15-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 2
-// CHECK15-NEXT:    store i64 [[TMP68]], i64* [[TMP84]], align 4
-// CHECK15-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
-// CHECK15-NEXT:    store i8* null, i8** [[TMP85]], align 4
-// CHECK15-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK15-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK15-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP67]])
+// CHECK15-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK15-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK15-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK15-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK15-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK15-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK15-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK15-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK15-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK15-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK15-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK15:       omp_offload.failed16:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK15:       omp_offload.cont17:
+// CHECK15-NEXT:    [[TMP78:%.*]] = load i32, i32* [[M]], align 4
+// CHECK15-NEXT:    store i32 [[TMP78]], i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK15-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP79]], i32* [[N_CASTED19]], align 4
+// CHECK15-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_CASTED19]], align 4
+// CHECK15-NEXT:    [[TMP81:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK15-NEXT:    store i32 [[TMP81]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK15-NEXT:    [[TMP82:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK15-NEXT:    [[TMP83:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK15-NEXT:    [[TMP84:%.*]] = sext i32 [[TMP83]] to i64
+// CHECK15-NEXT:    [[TMP85:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES23]] to i8*
+// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP85]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
+// CHECK15-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
 // CHECK15-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP66]], i32* [[TMP87]], align 4
-// CHECK15-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK15-NEXT:    store i32 [[TMP80]], i32* [[TMP87]], align 4
+// CHECK15-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
 // CHECK15-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP66]], i32* [[TMP89]], align 4
-// CHECK15-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK15-NEXT:    store i32 [[TMP80]], i32* [[TMP89]], align 4
+// CHECK15-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
 // CHECK15-NEXT:    store i8* null, i8** [[TMP90]], align 4
-// CHECK15-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP94:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP94]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK15-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK15-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP95]], 0
-// CHECK15-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
-// CHECK15-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
-// CHECK15-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK15-NEXT:    [[TMP96:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK15-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP96]], 1
-// CHECK15-NEXT:    [[TMP97:%.*]] = zext i32 [[ADD29]] to i64
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP97]])
-// CHECK15-NEXT:    [[TMP98:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, i32 4, i8** [[TMP91]], i8** [[TMP92]], i64* [[TMP93]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP99:%.*]] = icmp ne i32 [[TMP98]], 0
-// CHECK15-NEXT:    br i1 [[TMP99]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK15:       omp_offload.failed30:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i32 [[TMP64]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP66]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK15:       omp_offload.cont31:
-// CHECK15-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP100]], i32* [[N_CASTED32]], align 4
-// CHECK15-NEXT:    [[TMP101:%.*]] = load i32, i32* [[N_CASTED32]], align 4
-// CHECK15-NEXT:    [[TMP102:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK15-NEXT:    [[TMP103:%.*]] = sext i32 [[TMP102]] to i64
-// CHECK15-NEXT:    [[TMP104:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES36]] to i8*
-// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP104]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
-// CHECK15-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP106:%.*]] = bitcast i8** [[TMP105]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP101]], i32* [[TMP106]], align 4
-// CHECK15-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP101]], i32* [[TMP108]], align 4
-// CHECK15-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP109]], align 4
-// CHECK15-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP111:%.*]] = bitcast i8** [[TMP110]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP111]], align 4
-// CHECK15-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP113]], align 4
-// CHECK15-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP114]], align 4
-// CHECK15-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP116]], align 4
-// CHECK15-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP118]], align 4
-// CHECK15-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES36]], i32 0, i32 2
-// CHECK15-NEXT:    store i64 [[TMP103]], i64* [[TMP119]], align 4
-// CHECK15-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 2
-// CHECK15-NEXT:    store i8* null, i8** [[TMP120]], align 4
-// CHECK15-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES36]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP92]], align 4
+// CHECK15-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP94]], align 4
+// CHECK15-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP95]], align 4
+// CHECK15-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 4
+// CHECK15-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 4
+// CHECK15-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 2
+// CHECK15-NEXT:    store i64 [[TMP84]], i64* [[TMP100]], align 4
+// CHECK15-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 2
+// CHECK15-NEXT:    store i8* null, i8** [[TMP101]], align 4
+// CHECK15-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP82]], i32* [[TMP103]], align 4
+// CHECK15-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP82]], i32* [[TMP105]], align 4
+// CHECK15-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 3
+// CHECK15-NEXT:    store i8* null, i8** [[TMP106]], align 4
+// CHECK15-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP110:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP110]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK15-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK15-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP111]], 0
+// CHECK15-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
+// CHECK15-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
+// CHECK15-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK15-NEXT:    [[TMP112:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK15-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP112]], 1
+// CHECK15-NEXT:    [[TMP113:%.*]] = zext i32 [[ADD30]] to i64
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP113]])
+// CHECK15-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP114]], align 4
+// CHECK15-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 4, i32* [[TMP115]], align 4
+// CHECK15-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK15-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP108]], i8*** [[TMP117]], align 4
+// CHECK15-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP109]], i64** [[TMP118]], align 4
+// CHECK15-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP119]], align 4
+// CHECK15-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK15-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP121]], align 4
+// CHECK15-NEXT:    [[TMP122:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK15-NEXT:    [[TMP123:%.*]] = icmp ne i32 [[TMP122]], 0
+// CHECK15-NEXT:    br i1 [[TMP123]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK15:       omp_offload.failed32:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i32 [[TMP80]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP82]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK15:       omp_offload.cont33:
 // CHECK15-NEXT:    [[TMP124:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP124]], i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK15-NEXT:    [[TMP125:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK15-NEXT:    [[SUB40:%.*]] = sub nsw i32 [[TMP125]], 0
-// CHECK15-NEXT:    [[DIV41:%.*]] = sdiv i32 [[SUB40]], 1
-// CHECK15-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[DIV41]], 1
-// CHECK15-NEXT:    store i32 [[SUB42]], i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK15-NEXT:    [[TMP126:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK15-NEXT:    [[ADD43:%.*]] = add nsw i32 [[TMP126]], 1
-// CHECK15-NEXT:    [[TMP127:%.*]] = zext i32 [[ADD43]] to i64
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP127]])
-// CHECK15-NEXT:    [[TMP128:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, i32 3, i8** [[TMP121]], i8** [[TMP122]], i64* [[TMP123]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP129:%.*]] = icmp ne i32 [[TMP128]], 0
-// CHECK15-NEXT:    br i1 [[TMP129]], label [[OMP_OFFLOAD_FAILED44:%.*]], label [[OMP_OFFLOAD_CONT45:%.*]]
-// CHECK15:       omp_offload.failed44:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i32 [[TMP101]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT45]]
-// CHECK15:       omp_offload.cont45:
-// CHECK15-NEXT:    [[TMP130:%.*]] = load i32, i32* [[M]], align 4
-// CHECK15-NEXT:    store i32 [[TMP130]], i32* [[DOTCAPTURE_EXPR_46]], align 4
-// CHECK15-NEXT:    [[TMP131:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP131]], i32* [[N_CASTED47]], align 4
-// CHECK15-NEXT:    [[TMP132:%.*]] = load i32, i32* [[N_CASTED47]], align 4
-// CHECK15-NEXT:    [[TMP133:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_46]], align 4
-// CHECK15-NEXT:    store i32 [[TMP133]], i32* [[DOTCAPTURE_EXPR__CASTED48]], align 4
-// CHECK15-NEXT:    [[TMP134:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED48]], align 4
-// CHECK15-NEXT:    [[TMP135:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK15-NEXT:    [[TMP136:%.*]] = sext i32 [[TMP135]] to i64
-// CHECK15-NEXT:    [[TMP137:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES52]] to i8*
-// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP137]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
-// CHECK15-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP132]], i32* [[TMP139]], align 4
-// CHECK15-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP132]], i32* [[TMP141]], align 4
-// CHECK15-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP142]], align 4
-// CHECK15-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP144]], align 4
-// CHECK15-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP146]], align 4
-// CHECK15-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP147]], align 4
-// CHECK15-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP149]], align 4
-// CHECK15-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP151]], align 4
-// CHECK15-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES52]], i32 0, i32 2
-// CHECK15-NEXT:    store i64 [[TMP136]], i64* [[TMP152]], align 4
-// CHECK15-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 2
-// CHECK15-NEXT:    store i8* null, i8** [[TMP153]], align 4
-// CHECK15-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 3
-// CHECK15-NEXT:    [[TMP155:%.*]] = bitcast i8** [[TMP154]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP134]], i32* [[TMP155]], align 4
-// CHECK15-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 3
-// CHECK15-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP134]], i32* [[TMP157]], align 4
-// CHECK15-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 3
-// CHECK15-NEXT:    store i8* null, i8** [[TMP158]], align 4
-// CHECK15-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES52]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP162:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP162]], i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK15-NEXT:    [[TMP163:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK15-NEXT:    [[SUB56:%.*]] = sub nsw i32 [[TMP163]], 0
-// CHECK15-NEXT:    [[DIV57:%.*]] = sdiv i32 [[SUB56]], 1
-// CHECK15-NEXT:    [[SUB58:%.*]] = sub nsw i32 [[DIV57]], 1
-// CHECK15-NEXT:    store i32 [[SUB58]], i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK15-NEXT:    [[TMP164:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK15-NEXT:    [[ADD59:%.*]] = add nsw i32 [[TMP164]], 1
-// CHECK15-NEXT:    [[TMP165:%.*]] = zext i32 [[ADD59]] to i64
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP165]])
-// CHECK15-NEXT:    [[TMP166:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, i32 4, i8** [[TMP159]], i8** [[TMP160]], i64* [[TMP161]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP167:%.*]] = icmp ne i32 [[TMP166]], 0
-// CHECK15-NEXT:    br i1 [[TMP167]], label [[OMP_OFFLOAD_FAILED60:%.*]], label [[OMP_OFFLOAD_CONT61:%.*]]
-// CHECK15:       omp_offload.failed60:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i32 [[TMP132]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP134]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT61]]
-// CHECK15:       omp_offload.cont61:
-// CHECK15-NEXT:    [[TMP168:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK15-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP168]])
+// CHECK15-NEXT:    store i32 [[TMP124]], i32* [[N_CASTED34]], align 4
+// CHECK15-NEXT:    [[TMP125:%.*]] = load i32, i32* [[N_CASTED34]], align 4
+// CHECK15-NEXT:    [[TMP126:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK15-NEXT:    [[TMP127:%.*]] = sext i32 [[TMP126]] to i64
+// CHECK15-NEXT:    [[TMP128:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES38]] to i8*
+// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP128]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
+// CHECK15-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP125]], i32* [[TMP130]], align 4
+// CHECK15-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP125]], i32* [[TMP132]], align 4
+// CHECK15-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS37]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP133]], align 4
+// CHECK15-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP135]], align 4
+// CHECK15-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP137]], align 4
+// CHECK15-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS37]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP138]], align 4
+// CHECK15-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP140]], align 4
+// CHECK15-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP142]], align 4
+// CHECK15-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES38]], i32 0, i32 2
+// CHECK15-NEXT:    store i64 [[TMP127]], i64* [[TMP143]], align 4
+// CHECK15-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS37]], i32 0, i32 2
+// CHECK15-NEXT:    store i8* null, i8** [[TMP144]], align 4
+// CHECK15-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES38]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP148:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP148]], i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK15-NEXT:    [[TMP149:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK15-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[TMP149]], 0
+// CHECK15-NEXT:    [[DIV43:%.*]] = sdiv i32 [[SUB42]], 1
+// CHECK15-NEXT:    [[SUB44:%.*]] = sub nsw i32 [[DIV43]], 1
+// CHECK15-NEXT:    store i32 [[SUB44]], i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK15-NEXT:    [[TMP150:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK15-NEXT:    [[ADD45:%.*]] = add nsw i32 [[TMP150]], 1
+// CHECK15-NEXT:    [[TMP151:%.*]] = zext i32 [[ADD45]] to i64
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP151]])
+// CHECK15-NEXT:    [[KERNEL_ARGS46:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP152]], align 4
+// CHECK15-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 3, i32* [[TMP153]], align 4
+// CHECK15-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP145]], i8*** [[TMP154]], align 4
+// CHECK15-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP146]], i8*** [[TMP155]], align 4
+// CHECK15-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP147]], i64** [[TMP156]], align 4
+// CHECK15-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP157]], align 4
+// CHECK15-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP158]], align 4
+// CHECK15-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP159]], align 4
+// CHECK15-NEXT:    [[TMP160:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]])
+// CHECK15-NEXT:    [[TMP161:%.*]] = icmp ne i32 [[TMP160]], 0
+// CHECK15-NEXT:    br i1 [[TMP161]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
+// CHECK15:       omp_offload.failed47:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i32 [[TMP125]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
+// CHECK15:       omp_offload.cont48:
+// CHECK15-NEXT:    [[TMP162:%.*]] = load i32, i32* [[M]], align 4
+// CHECK15-NEXT:    store i32 [[TMP162]], i32* [[DOTCAPTURE_EXPR_49]], align 4
+// CHECK15-NEXT:    [[TMP163:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP163]], i32* [[N_CASTED50]], align 4
+// CHECK15-NEXT:    [[TMP164:%.*]] = load i32, i32* [[N_CASTED50]], align 4
+// CHECK15-NEXT:    [[TMP165:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_49]], align 4
+// CHECK15-NEXT:    store i32 [[TMP165]], i32* [[DOTCAPTURE_EXPR__CASTED51]], align 4
+// CHECK15-NEXT:    [[TMP166:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED51]], align 4
+// CHECK15-NEXT:    [[TMP167:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK15-NEXT:    [[TMP168:%.*]] = sext i32 [[TMP167]] to i64
+// CHECK15-NEXT:    [[TMP169:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES55]] to i8*
+// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP169]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
+// CHECK15-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP171:%.*]] = bitcast i8** [[TMP170]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP164]], i32* [[TMP171]], align 4
+// CHECK15-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP173:%.*]] = bitcast i8** [[TMP172]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP164]], i32* [[TMP173]], align 4
+// CHECK15-NEXT:    [[TMP174:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP174]], align 4
+// CHECK15-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP176:%.*]] = bitcast i8** [[TMP175]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP176]], align 4
+// CHECK15-NEXT:    [[TMP177:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP178:%.*]] = bitcast i8** [[TMP177]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP178]], align 4
+// CHECK15-NEXT:    [[TMP179:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP179]], align 4
+// CHECK15-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP181]], align 4
+// CHECK15-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP183:%.*]] = bitcast i8** [[TMP182]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP183]], align 4
+// CHECK15-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES55]], i32 0, i32 2
+// CHECK15-NEXT:    store i64 [[TMP168]], i64* [[TMP184]], align 4
+// CHECK15-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 2
+// CHECK15-NEXT:    store i8* null, i8** [[TMP185]], align 4
+// CHECK15-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP187:%.*]] = bitcast i8** [[TMP186]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP166]], i32* [[TMP187]], align 4
+// CHECK15-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP189:%.*]] = bitcast i8** [[TMP188]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP166]], i32* [[TMP189]], align 4
+// CHECK15-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 3
+// CHECK15-NEXT:    store i8* null, i8** [[TMP190]], align 4
+// CHECK15-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES55]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP194:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP194]], i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK15-NEXT:    [[TMP195:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK15-NEXT:    [[SUB59:%.*]] = sub nsw i32 [[TMP195]], 0
+// CHECK15-NEXT:    [[DIV60:%.*]] = sdiv i32 [[SUB59]], 1
+// CHECK15-NEXT:    [[SUB61:%.*]] = sub nsw i32 [[DIV60]], 1
+// CHECK15-NEXT:    store i32 [[SUB61]], i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK15-NEXT:    [[TMP196:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK15-NEXT:    [[ADD62:%.*]] = add nsw i32 [[TMP196]], 1
+// CHECK15-NEXT:    [[TMP197:%.*]] = zext i32 [[ADD62]] to i64
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP197]])
+// CHECK15-NEXT:    [[KERNEL_ARGS63:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP198]], align 4
+// CHECK15-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 4, i32* [[TMP199]], align 4
+// CHECK15-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP191]], i8*** [[TMP200]], align 4
+// CHECK15-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP192]], i8*** [[TMP201]], align 4
+// CHECK15-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP193]], i64** [[TMP202]], align 4
+// CHECK15-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP203]], align 4
+// CHECK15-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP204]], align 4
+// CHECK15-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP205]], align 4
+// CHECK15-NEXT:    [[TMP206:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]])
+// CHECK15-NEXT:    [[TMP207:%.*]] = icmp ne i32 [[TMP206]], 0
+// CHECK15-NEXT:    br i1 [[TMP207]], label [[OMP_OFFLOAD_FAILED64:%.*]], label [[OMP_OFFLOAD_CONT65:%.*]]
+// CHECK15:       omp_offload.failed64:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i32 [[TMP164]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP166]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT65]]
+// CHECK15:       omp_offload.cont65:
+// CHECK15-NEXT:    [[TMP208:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK15-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP208]])
 // CHECK15-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK15-NEXT:    [[TMP169:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK15-NEXT:    call void @llvm.stackrestore(i8* [[TMP169]])
-// CHECK15-NEXT:    [[TMP170:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK15-NEXT:    ret i32 [[TMP170]]
+// CHECK15-NEXT:    [[TMP209:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK15-NEXT:    call void @llvm.stackrestore(i8* [[TMP209]])
+// CHECK15-NEXT:    [[TMP210:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK15-NEXT:    ret i32 [[TMP210]]
 //
 //
 // CHECK15-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139
@@ -7792,20 +8387,20 @@ int main (int argc, char **argv) {
 // CHECK15-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 4
-// CHECK15-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR__CASTED20:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 4
+// CHECK15-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR__CASTED23:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK15-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK15-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -7819,111 +8414,196 @@ int main (int argc, char **argv) {
 // CHECK15-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK15-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK15-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK15-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK15-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK15-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK15-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK15-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK15-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK15-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK15-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK15-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK15-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK15-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK15-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK15:       omp_offload.failed:
 // CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK15:       omp_offload.cont:
-// CHECK15-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK15-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK15-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK15-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK15-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK15-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK15-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK15-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK15:       omp_offload.failed5:
+// CHECK15-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK15-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK15-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK15-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK15-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK15-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK15-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK15-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK15-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK15-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK15-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK15:       omp_offload.failed6:
 // CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK15:       omp_offload.cont6:
-// CHECK15-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK15-NEXT:    store i32 [[TMP18]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK15-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK15-NEXT:    store i32 [[TMP19]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK15-NEXT:    [[TMP20:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK15-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP22]], align 4
-// CHECK15-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP24]], align 4
-// CHECK15-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP25]], align 4
-// CHECK15-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP20]], i32* [[TMP27]], align 4
-// CHECK15-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP20]], i32* [[TMP29]], align 4
-// CHECK15-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP30]], align 4
-// CHECK15-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, i32 2, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK15-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK15:       omp_offload.failed11:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i32 [[TMP20]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK15:       omp_offload.cont12:
-// CHECK15-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP36]], align 4
-// CHECK15-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK15:       omp_offload.cont7:
+// CHECK15-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK15-NEXT:    store i32 [[TMP34]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK15-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK15-NEXT:    store i32 [[TMP35]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK15-NEXT:    [[TMP36:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK15-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
 // CHECK15-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [10 x i32]**
 // CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP38]], align 4
-// CHECK15-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP39]], align 4
-// CHECK15-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP40]], align 4
+// CHECK15-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP41]], align 4
+// CHECK15-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP36]], i32* [[TMP43]], align 4
+// CHECK15-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP36]], i32* [[TMP45]], align 4
+// CHECK15-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP46]], align 4
+// CHECK15-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
+// CHECK15-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP49]], align 4
+// CHECK15-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 2, i32* [[TMP50]], align 4
+// CHECK15-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 4
+// CHECK15-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP48]], i8*** [[TMP52]], align 4
+// CHECK15-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP53]], align 4
+// CHECK15-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP54]], align 4
+// CHECK15-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP55]], align 4
+// CHECK15-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP56]], align 4
+// CHECK15-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK15-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
+// CHECK15-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK15:       omp_offload.failed13:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i32 [[TMP36]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK15:       omp_offload.cont14:
+// CHECK15-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP60]], align 4
+// CHECK15-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP62]], align 4
+// CHECK15-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP63]], align 4
+// CHECK15-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, i32 1, i8** [[TMP40]], i8** [[TMP41]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK15-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK15:       omp_offload.failed17:
+// CHECK15-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK15-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK15-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK15-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 4
+// CHECK15-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK15-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP71]], align 4
+// CHECK15-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK15-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP73]], align 4
+// CHECK15-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK15-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK15-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK15:       omp_offload.failed20:
 // CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK15:       omp_offload.cont18:
-// CHECK15-NEXT:    [[TMP44:%.*]] = load i32, i32* [[M]], align 4
-// CHECK15-NEXT:    store i32 [[TMP44]], i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK15-NEXT:    [[TMP45:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK15-NEXT:    store i32 [[TMP45]], i32* [[DOTCAPTURE_EXPR__CASTED20]], align 4
-// CHECK15-NEXT:    [[TMP46:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED20]], align 4
-// CHECK15-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP48]], align 4
-// CHECK15-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP50]], align 4
-// CHECK15-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP51]], align 4
-// CHECK15-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP46]], i32* [[TMP53]], align 4
-// CHECK15-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP46]], i32* [[TMP55]], align 4
-// CHECK15-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP56]], align 4
-// CHECK15-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK15:       omp_offload.cont21:
+// CHECK15-NEXT:    [[TMP76:%.*]] = load i32, i32* [[M]], align 4
+// CHECK15-NEXT:    store i32 [[TMP76]], i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK15-NEXT:    [[TMP77:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK15-NEXT:    store i32 [[TMP77]], i32* [[DOTCAPTURE_EXPR__CASTED23]], align 4
+// CHECK15-NEXT:    [[TMP78:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED23]], align 4
+// CHECK15-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP80]], align 4
+// CHECK15-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP82]], align 4
+// CHECK15-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP83]], align 4
+// CHECK15-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP78]], i32* [[TMP85]], align 4
+// CHECK15-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP78]], i32* [[TMP87]], align 4
+// CHECK15-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP88]], align 4
+// CHECK15-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, i32 2, i8** [[TMP57]], i8** [[TMP58]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK15-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
-// CHECK15:       omp_offload.failed25:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i32 [[TMP46]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
-// CHECK15:       omp_offload.cont26:
+// CHECK15-NEXT:    [[KERNEL_ARGS28:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP91]], align 4
+// CHECK15-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 2, i32* [[TMP92]], align 4
+// CHECK15-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP89]], i8*** [[TMP93]], align 4
+// CHECK15-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP90]], i8*** [[TMP94]], align 4
+// CHECK15-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP95]], align 4
+// CHECK15-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP96]], align 4
+// CHECK15-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP97]], align 4
+// CHECK15-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP98]], align 4
+// CHECK15-NEXT:    [[TMP99:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]])
+// CHECK15-NEXT:    [[TMP100:%.*]] = icmp ne i32 [[TMP99]], 0
+// CHECK15-NEXT:    br i1 [[TMP100]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
+// CHECK15:       omp_offload.failed29:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i32 [[TMP78]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
+// CHECK15:       omp_offload.cont30:
 // CHECK15-NEXT:    ret i32 0
 //
 //
@@ -8730,34 +9410,34 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[_TMP9:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
-// CHECK17-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES26:%.*]] = alloca [4 x i64], align 8
+// CHECK17-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[N_CASTED35:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS37:%.*]] = alloca [3 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS38:%.*]] = alloca [3 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS39:%.*]] = alloca [3 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES40:%.*]] = alloca [3 x i64], align 8
-// CHECK17-NEXT:    [[_TMP41:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_42:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_43:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_50:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[N_CASTED51:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED53:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS55:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS56:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS57:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES58:%.*]] = alloca [4 x i64], align 8
-// CHECK17-NEXT:    [[_TMP59:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_60:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_61:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_29:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[N_CASTED37:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS39:%.*]] = alloca [3 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS40:%.*]] = alloca [3 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS41:%.*]] = alloca [3 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES42:%.*]] = alloca [3 x i64], align 8
+// CHECK17-NEXT:    [[_TMP43:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_44:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_45:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_53:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[N_CASTED54:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED56:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS58:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS59:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS60:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES61:%.*]] = alloca [4 x i64], align 8
+// CHECK17-NEXT:    [[_TMP62:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_63:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_64:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK17-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK17-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -8816,264 +9496,349 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK17-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK17-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK17-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK17-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK17-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK17-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK17-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK17-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK17-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK17-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
-// CHECK17-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK17-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK17-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK17-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK17-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
-// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK17-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK17-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK17-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK17-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK17-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK17-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK17-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK17-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK17-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK17-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK17-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK17-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK17-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
+// CHECK17-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK17-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK17-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK17-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK17-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK17-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK17-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK17-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK17-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK17-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK17-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK17-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK17-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK17-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK17-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK17-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK17-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP58]])
-// CHECK17-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK17-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK17:       omp_offload.failed16:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK17:       omp_offload.cont17:
-// CHECK17-NEXT:    [[TMP61:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    store i32 [[TMP61]], i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK17-NEXT:    [[TMP62:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP62]], i32* [[CONV20]], align 4
-// CHECK17-NEXT:    [[TMP63:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK17-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK17-NEXT:    [[CONV21:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP64]], i32* [[CONV21]], align 4
-// CHECK17-NEXT:    [[TMP65:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK17-NEXT:    [[TMP66:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK17-NEXT:    [[TMP67:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP67]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
-// CHECK17-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP63]], i64* [[TMP69]], align 8
-// CHECK17-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP63]], i64* [[TMP71]], align 8
-// CHECK17-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP72]], align 8
-// CHECK17-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP74]], align 8
-// CHECK17-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP76]], align 8
-// CHECK17-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP77]], align 8
-// CHECK17-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP79]], align 8
-// CHECK17-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 8
-// CHECK17-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 2
-// CHECK17-NEXT:    store i64 [[TMP66]], i64* [[TMP82]], align 8
-// CHECK17-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP83]], align 8
-// CHECK17-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK17-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK17-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP66]])
+// CHECK17-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK17-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK17-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK17-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK17-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK17-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK17-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK17-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK17-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK17-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK17-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK17:       omp_offload.failed17:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK17:       omp_offload.cont18:
+// CHECK17-NEXT:    [[TMP77:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    store i32 [[TMP77]], i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK17-NEXT:    [[TMP78:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP78]], i32* [[CONV21]], align 4
+// CHECK17-NEXT:    [[TMP79:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK17-NEXT:    [[TMP80:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK17-NEXT:    [[CONV22:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP80]], i32* [[CONV22]], align 4
+// CHECK17-NEXT:    [[TMP81:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK17-NEXT:    [[TMP82:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK17-NEXT:    [[TMP83:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES26]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP83]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
+// CHECK17-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP65]], i64* [[TMP85]], align 8
-// CHECK17-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK17-NEXT:    store i64 [[TMP79]], i64* [[TMP85]], align 8
+// CHECK17-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP65]], i64* [[TMP87]], align 8
-// CHECK17-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK17-NEXT:    store i64 [[TMP79]], i64* [[TMP87]], align 8
+// CHECK17-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 0
 // CHECK17-NEXT:    store i8* null, i8** [[TMP88]], align 8
-// CHECK17-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP92:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP92]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK17-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK17-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP93]], 0
-// CHECK17-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
-// CHECK17-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
-// CHECK17-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK17-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK17-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP94]], 1
-// CHECK17-NEXT:    [[TMP95:%.*]] = zext i32 [[ADD32]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP95]])
-// CHECK17-NEXT:    [[TMP96:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, i32 4, i8** [[TMP89]], i8** [[TMP90]], i64* [[TMP91]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP97:%.*]] = icmp ne i32 [[TMP96]], 0
-// CHECK17-NEXT:    br i1 [[TMP97]], label [[OMP_OFFLOAD_FAILED33:%.*]], label [[OMP_OFFLOAD_CONT34:%.*]]
-// CHECK17:       omp_offload.failed33:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i64 [[TMP63]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP65]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT34]]
-// CHECK17:       omp_offload.cont34:
-// CHECK17-NEXT:    [[TMP98:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    [[CONV36:%.*]] = bitcast i64* [[N_CASTED35]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP98]], i32* [[CONV36]], align 4
-// CHECK17-NEXT:    [[TMP99:%.*]] = load i64, i64* [[N_CASTED35]], align 8
-// CHECK17-NEXT:    [[TMP100:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK17-NEXT:    [[TMP101:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES40]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP101]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
-// CHECK17-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP90]], align 8
+// CHECK17-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP92]], align 8
+// CHECK17-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP93]], align 8
+// CHECK17-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP95]], align 8
+// CHECK17-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 8
+// CHECK17-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 2
+// CHECK17-NEXT:    store i64 [[TMP82]], i64* [[TMP98]], align 8
+// CHECK17-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP99]], align 8
+// CHECK17-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP81]], i64* [[TMP101]], align 8
+// CHECK17-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 3
 // CHECK17-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP99]], i64* [[TMP103]], align 8
-// CHECK17-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP99]], i64* [[TMP105]], align 8
-// CHECK17-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS39]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP106]], align 8
-// CHECK17-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP108]], align 8
-// CHECK17-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP110]], align 8
-// CHECK17-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS39]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP111]], align 8
-// CHECK17-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP113]], align 8
-// CHECK17-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP115]], align 8
-// CHECK17-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES40]], i32 0, i32 2
-// CHECK17-NEXT:    store i64 [[TMP100]], i64* [[TMP116]], align 8
-// CHECK17-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS39]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP117]], align 8
-// CHECK17-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES40]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP121:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP121]], i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK17-NEXT:    [[TMP122:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK17-NEXT:    [[SUB44:%.*]] = sub nsw i32 [[TMP122]], 0
-// CHECK17-NEXT:    [[DIV45:%.*]] = sdiv i32 [[SUB44]], 1
-// CHECK17-NEXT:    [[SUB46:%.*]] = sub nsw i32 [[DIV45]], 1
-// CHECK17-NEXT:    store i32 [[SUB46]], i32* [[DOTCAPTURE_EXPR_43]], align 4
-// CHECK17-NEXT:    [[TMP123:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_43]], align 4
-// CHECK17-NEXT:    [[ADD47:%.*]] = add nsw i32 [[TMP123]], 1
-// CHECK17-NEXT:    [[TMP124:%.*]] = zext i32 [[ADD47]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP124]])
-// CHECK17-NEXT:    [[TMP125:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, i32 3, i8** [[TMP118]], i8** [[TMP119]], i64* [[TMP120]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP126:%.*]] = icmp ne i32 [[TMP125]], 0
-// CHECK17-NEXT:    br i1 [[TMP126]], label [[OMP_OFFLOAD_FAILED48:%.*]], label [[OMP_OFFLOAD_CONT49:%.*]]
-// CHECK17:       omp_offload.failed48:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i64 [[TMP99]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT49]]
-// CHECK17:       omp_offload.cont49:
-// CHECK17-NEXT:    [[TMP127:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    store i32 [[TMP127]], i32* [[DOTCAPTURE_EXPR_50]], align 4
-// CHECK17-NEXT:    [[TMP128:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    [[CONV52:%.*]] = bitcast i64* [[N_CASTED51]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP128]], i32* [[CONV52]], align 4
-// CHECK17-NEXT:    [[TMP129:%.*]] = load i64, i64* [[N_CASTED51]], align 8
-// CHECK17-NEXT:    [[TMP130:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_50]], align 4
-// CHECK17-NEXT:    [[CONV54:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED53]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP130]], i32* [[CONV54]], align 4
-// CHECK17-NEXT:    [[TMP131:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED53]], align 8
-// CHECK17-NEXT:    [[TMP132:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK17-NEXT:    [[TMP133:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES58]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP133]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
-// CHECK17-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP129]], i64* [[TMP135]], align 8
-// CHECK17-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP129]], i64* [[TMP137]], align 8
-// CHECK17-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP138]], align 8
-// CHECK17-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP140]], align 8
-// CHECK17-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP142]], align 8
-// CHECK17-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP143]], align 8
-// CHECK17-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP145:%.*]] = bitcast i8** [[TMP144]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP145]], align 8
-// CHECK17-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP147]], align 8
-// CHECK17-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES58]], i32 0, i32 2
-// CHECK17-NEXT:    store i64 [[TMP132]], i64* [[TMP148]], align 8
-// CHECK17-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP149]], align 8
-// CHECK17-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP131]], i64* [[TMP151]], align 8
-// CHECK17-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP131]], i64* [[TMP153]], align 8
-// CHECK17-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 3
-// CHECK17-NEXT:    store i8* null, i8** [[TMP154]], align 8
-// CHECK17-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES58]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP158:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP158]], i32* [[DOTCAPTURE_EXPR_60]], align 4
-// CHECK17-NEXT:    [[TMP159:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_60]], align 4
-// CHECK17-NEXT:    [[SUB62:%.*]] = sub nsw i32 [[TMP159]], 0
-// CHECK17-NEXT:    [[DIV63:%.*]] = sdiv i32 [[SUB62]], 1
-// CHECK17-NEXT:    [[SUB64:%.*]] = sub nsw i32 [[DIV63]], 1
-// CHECK17-NEXT:    store i32 [[SUB64]], i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK17-NEXT:    [[TMP160:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK17-NEXT:    [[ADD65:%.*]] = add nsw i32 [[TMP160]], 1
-// CHECK17-NEXT:    [[TMP161:%.*]] = zext i32 [[ADD65]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP161]])
-// CHECK17-NEXT:    [[TMP162:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, i32 4, i8** [[TMP155]], i8** [[TMP156]], i64* [[TMP157]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP163:%.*]] = icmp ne i32 [[TMP162]], 0
-// CHECK17-NEXT:    br i1 [[TMP163]], label [[OMP_OFFLOAD_FAILED66:%.*]], label [[OMP_OFFLOAD_CONT67:%.*]]
-// CHECK17:       omp_offload.failed66:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i64 [[TMP129]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP131]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT67]]
-// CHECK17:       omp_offload.cont67:
-// CHECK17-NEXT:    [[TMP164:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK17-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP164]])
+// CHECK17-NEXT:    store i64 [[TMP81]], i64* [[TMP103]], align 8
+// CHECK17-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 3
+// CHECK17-NEXT:    store i8* null, i8** [[TMP104]], align 8
+// CHECK17-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP108:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP108]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK17-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK17-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[TMP109]], 0
+// CHECK17-NEXT:    [[DIV31:%.*]] = sdiv i32 [[SUB30]], 1
+// CHECK17-NEXT:    [[SUB32:%.*]] = sub nsw i32 [[DIV31]], 1
+// CHECK17-NEXT:    store i32 [[SUB32]], i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK17-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK17-NEXT:    [[ADD33:%.*]] = add nsw i32 [[TMP110]], 1
+// CHECK17-NEXT:    [[TMP111:%.*]] = zext i32 [[ADD33]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP111]])
+// CHECK17-NEXT:    [[KERNEL_ARGS34:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP112]], align 4
+// CHECK17-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 4, i32* [[TMP113]], align 4
+// CHECK17-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK17-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 8
+// CHECK17-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP107]], i64** [[TMP116]], align 8
+// CHECK17-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP117]], align 8
+// CHECK17-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK17-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP119]], align 8
+// CHECK17-NEXT:    [[TMP120:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]])
+// CHECK17-NEXT:    [[TMP121:%.*]] = icmp ne i32 [[TMP120]], 0
+// CHECK17-NEXT:    br i1 [[TMP121]], label [[OMP_OFFLOAD_FAILED35:%.*]], label [[OMP_OFFLOAD_CONT36:%.*]]
+// CHECK17:       omp_offload.failed35:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i64 [[TMP79]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP81]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT36]]
+// CHECK17:       omp_offload.cont36:
+// CHECK17-NEXT:    [[TMP122:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    [[CONV38:%.*]] = bitcast i64* [[N_CASTED37]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP122]], i32* [[CONV38]], align 4
+// CHECK17-NEXT:    [[TMP123:%.*]] = load i64, i64* [[N_CASTED37]], align 8
+// CHECK17-NEXT:    [[TMP124:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK17-NEXT:    [[TMP125:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES42]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP125]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
+// CHECK17-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP123]], i64* [[TMP127]], align 8
+// CHECK17-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP123]], i64* [[TMP129]], align 8
+// CHECK17-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS41]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP130]], align 8
+// CHECK17-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP132]], align 8
+// CHECK17-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP134]], align 8
+// CHECK17-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS41]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP135]], align 8
+// CHECK17-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP137]], align 8
+// CHECK17-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP139]], align 8
+// CHECK17-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES42]], i32 0, i32 2
+// CHECK17-NEXT:    store i64 [[TMP124]], i64* [[TMP140]], align 8
+// CHECK17-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS41]], i64 0, i64 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP141]], align 8
+// CHECK17-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES42]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP145:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP145]], i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK17-NEXT:    [[TMP146:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK17-NEXT:    [[SUB46:%.*]] = sub nsw i32 [[TMP146]], 0
+// CHECK17-NEXT:    [[DIV47:%.*]] = sdiv i32 [[SUB46]], 1
+// CHECK17-NEXT:    [[SUB48:%.*]] = sub nsw i32 [[DIV47]], 1
+// CHECK17-NEXT:    store i32 [[SUB48]], i32* [[DOTCAPTURE_EXPR_45]], align 4
+// CHECK17-NEXT:    [[TMP147:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_45]], align 4
+// CHECK17-NEXT:    [[ADD49:%.*]] = add nsw i32 [[TMP147]], 1
+// CHECK17-NEXT:    [[TMP148:%.*]] = zext i32 [[ADD49]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP148]])
+// CHECK17-NEXT:    [[KERNEL_ARGS50:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP149]], align 4
+// CHECK17-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP150]], align 4
+// CHECK17-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP142]], i8*** [[TMP151]], align 8
+// CHECK17-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP143]], i8*** [[TMP152]], align 8
+// CHECK17-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP144]], i64** [[TMP153]], align 8
+// CHECK17-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP154]], align 8
+// CHECK17-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP155]], align 8
+// CHECK17-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP156]], align 8
+// CHECK17-NEXT:    [[TMP157:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]])
+// CHECK17-NEXT:    [[TMP158:%.*]] = icmp ne i32 [[TMP157]], 0
+// CHECK17-NEXT:    br i1 [[TMP158]], label [[OMP_OFFLOAD_FAILED51:%.*]], label [[OMP_OFFLOAD_CONT52:%.*]]
+// CHECK17:       omp_offload.failed51:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i64 [[TMP123]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT52]]
+// CHECK17:       omp_offload.cont52:
+// CHECK17-NEXT:    [[TMP159:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    store i32 [[TMP159]], i32* [[DOTCAPTURE_EXPR_53]], align 4
+// CHECK17-NEXT:    [[TMP160:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    [[CONV55:%.*]] = bitcast i64* [[N_CASTED54]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP160]], i32* [[CONV55]], align 4
+// CHECK17-NEXT:    [[TMP161:%.*]] = load i64, i64* [[N_CASTED54]], align 8
+// CHECK17-NEXT:    [[TMP162:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_53]], align 4
+// CHECK17-NEXT:    [[CONV57:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED56]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP162]], i32* [[CONV57]], align 4
+// CHECK17-NEXT:    [[TMP163:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED56]], align 8
+// CHECK17-NEXT:    [[TMP164:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK17-NEXT:    [[TMP165:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES61]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP165]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
+// CHECK17-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP161]], i64* [[TMP167]], align 8
+// CHECK17-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP161]], i64* [[TMP169]], align 8
+// CHECK17-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP170]], align 8
+// CHECK17-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP172:%.*]] = bitcast i8** [[TMP171]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP172]], align 8
+// CHECK17-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP174:%.*]] = bitcast i8** [[TMP173]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP174]], align 8
+// CHECK17-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP175]], align 8
+// CHECK17-NEXT:    [[TMP176:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP177:%.*]] = bitcast i8** [[TMP176]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP177]], align 8
+// CHECK17-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP179]], align 8
+// CHECK17-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES61]], i32 0, i32 2
+// CHECK17-NEXT:    store i64 [[TMP164]], i64* [[TMP180]], align 8
+// CHECK17-NEXT:    [[TMP181:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP181]], align 8
+// CHECK17-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP183:%.*]] = bitcast i8** [[TMP182]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP163]], i64* [[TMP183]], align 8
+// CHECK17-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP185:%.*]] = bitcast i8** [[TMP184]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP163]], i64* [[TMP185]], align 8
+// CHECK17-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 3
+// CHECK17-NEXT:    store i8* null, i8** [[TMP186]], align 8
+// CHECK17-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES61]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP190:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP190]], i32* [[DOTCAPTURE_EXPR_63]], align 4
+// CHECK17-NEXT:    [[TMP191:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_63]], align 4
+// CHECK17-NEXT:    [[SUB65:%.*]] = sub nsw i32 [[TMP191]], 0
+// CHECK17-NEXT:    [[DIV66:%.*]] = sdiv i32 [[SUB65]], 1
+// CHECK17-NEXT:    [[SUB67:%.*]] = sub nsw i32 [[DIV66]], 1
+// CHECK17-NEXT:    store i32 [[SUB67]], i32* [[DOTCAPTURE_EXPR_64]], align 4
+// CHECK17-NEXT:    [[TMP192:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_64]], align 4
+// CHECK17-NEXT:    [[ADD68:%.*]] = add nsw i32 [[TMP192]], 1
+// CHECK17-NEXT:    [[TMP193:%.*]] = zext i32 [[ADD68]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP193]])
+// CHECK17-NEXT:    [[KERNEL_ARGS69:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP194]], align 4
+// CHECK17-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 4, i32* [[TMP195]], align 4
+// CHECK17-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP187]], i8*** [[TMP196]], align 8
+// CHECK17-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP188]], i8*** [[TMP197]], align 8
+// CHECK17-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP189]], i64** [[TMP198]], align 8
+// CHECK17-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP199]], align 8
+// CHECK17-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP200]], align 8
+// CHECK17-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP201]], align 8
+// CHECK17-NEXT:    [[TMP202:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]])
+// CHECK17-NEXT:    [[TMP203:%.*]] = icmp ne i32 [[TMP202]], 0
+// CHECK17-NEXT:    br i1 [[TMP203]], label [[OMP_OFFLOAD_FAILED70:%.*]], label [[OMP_OFFLOAD_CONT71:%.*]]
+// CHECK17:       omp_offload.failed70:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i64 [[TMP161]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP163]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT71]]
+// CHECK17:       omp_offload.cont71:
+// CHECK17-NEXT:    [[TMP204:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK17-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP204]])
 // CHECK17-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK17-NEXT:    [[TMP165:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP165]])
-// CHECK17-NEXT:    [[TMP166:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK17-NEXT:    ret i32 [[TMP166]]
+// CHECK17-NEXT:    [[TMP205:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP205]])
+// CHECK17-NEXT:    [[TMP206:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK17-NEXT:    ret i32 [[TMP206]]
 //
 //
 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139
@@ -10258,20 +11023,20 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 8
-// CHECK17-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED20:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 8
+// CHECK17-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED23:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS25:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS26:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS27:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[_TMP28:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK17-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK17-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -10285,113 +11050,198 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK17-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK17-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK17-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK17-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK17-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK17-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK17-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK17-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
-// CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK17-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK17-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK17-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK17-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK17-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK17:       omp_offload.failed5:
+// CHECK17-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK17-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK17-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK17-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK17-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK17-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK17:       omp_offload.failed6:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK17:       omp_offload.cont6:
-// CHECK17-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    store i32 [[TMP18]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK17-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK17:       omp_offload.cont7:
+// CHECK17-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    store i32 [[TMP34]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK17-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP19]], i32* [[CONV]], align 4
-// CHECK17-NEXT:    [[TMP20:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP22]], align 8
-// CHECK17-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP24]], align 8
-// CHECK17-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP25]], align 8
-// CHECK17-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP20]], i64* [[TMP27]], align 8
-// CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP20]], i64* [[TMP29]], align 8
-// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP30]], align 8
-// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, i32 2, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK17-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK17:       omp_offload.failed11:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i64 [[TMP20]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK17:       omp_offload.cont12:
-// CHECK17-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP36]], align 8
-// CHECK17-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 [[TMP35]], i32* [[CONV]], align 4
+// CHECK17-NEXT:    [[TMP36:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK17-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [10 x i32]**
 // CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP38]], align 8
-// CHECK17-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP39]], align 8
-// CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP40]], align 8
+// CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP41]], align 8
+// CHECK17-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP36]], i64* [[TMP43]], align 8
+// CHECK17-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP36]], i64* [[TMP45]], align 8
+// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP46]], align 8
+// CHECK17-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, i32 1, i8** [[TMP40]], i8** [[TMP41]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK17-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK17:       omp_offload.failed17:
+// CHECK17-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP49]], align 4
+// CHECK17-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 2, i32* [[TMP50]], align 4
+// CHECK17-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 8
+// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP48]], i8*** [[TMP52]], align 8
+// CHECK17-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP54]], align 8
+// CHECK17-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK17-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP56]], align 8
+// CHECK17-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK17-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
+// CHECK17-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK17:       omp_offload.failed13:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i64 [[TMP36]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK17:       omp_offload.cont14:
+// CHECK17-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP60]], align 8
+// CHECK17-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP62]], align 8
+// CHECK17-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP63]], align 8
+// CHECK17-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
+// CHECK17-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK17-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK17-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 8
+// CHECK17-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 8
+// CHECK17-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP70]], align 8
+// CHECK17-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP71]], align 8
+// CHECK17-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP72]], align 8
+// CHECK17-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK17-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK17-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK17-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK17:       omp_offload.failed20:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK17:       omp_offload.cont18:
-// CHECK17-NEXT:    [[TMP44:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    store i32 [[TMP44]], i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK17-NEXT:    [[TMP45:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK17-NEXT:    [[CONV21:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED20]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP45]], i32* [[CONV21]], align 4
-// CHECK17-NEXT:    [[TMP46:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED20]], align 8
-// CHECK17-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP48]], align 8
-// CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP50]], align 8
-// CHECK17-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP46]], i64* [[TMP53]], align 8
-// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP46]], i64* [[TMP55]], align 8
-// CHECK17-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP56]], align 8
-// CHECK17-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK17:       omp_offload.cont21:
+// CHECK17-NEXT:    [[TMP76:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    store i32 [[TMP76]], i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK17-NEXT:    [[TMP77:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK17-NEXT:    [[CONV24:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED23]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP77]], i32* [[CONV24]], align 4
+// CHECK17-NEXT:    [[TMP78:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED23]], align 8
+// CHECK17-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP80]], align 8
+// CHECK17-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP82]], align 8
+// CHECK17-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP83]], align 8
+// CHECK17-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP78]], i64* [[TMP85]], align 8
+// CHECK17-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP78]], i64* [[TMP87]], align 8
+// CHECK17-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP88]], align 8
+// CHECK17-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, i32 2, i8** [[TMP57]], i8** [[TMP58]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK17-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED26:%.*]], label [[OMP_OFFLOAD_CONT27:%.*]]
-// CHECK17:       omp_offload.failed26:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i64 [[TMP46]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT27]]
-// CHECK17:       omp_offload.cont27:
+// CHECK17-NEXT:    [[KERNEL_ARGS29:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP91]], align 4
+// CHECK17-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 2, i32* [[TMP92]], align 4
+// CHECK17-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP89]], i8*** [[TMP93]], align 8
+// CHECK17-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP90]], i8*** [[TMP94]], align 8
+// CHECK17-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP95]], align 8
+// CHECK17-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP96]], align 8
+// CHECK17-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP97]], align 8
+// CHECK17-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP98]], align 8
+// CHECK17-NEXT:    [[TMP99:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]])
+// CHECK17-NEXT:    [[TMP100:%.*]] = icmp ne i32 [[TMP99]], 0
+// CHECK17-NEXT:    br i1 [[TMP100]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
+// CHECK17:       omp_offload.failed30:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i64 [[TMP78]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
+// CHECK17:       omp_offload.cont31:
 // CHECK17-NEXT:    ret i32 0
 //
 //
@@ -11235,34 +12085,34 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[_TMP8:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_17:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[N_CASTED19:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
-// CHECK19-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES23:%.*]] = alloca [4 x i64], align 4
+// CHECK19-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[N_CASTED32:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [3 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [3 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [3 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES36:%.*]] = alloca [3 x i64], align 4
-// CHECK19-NEXT:    [[_TMP37:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_38:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_39:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_46:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[N_CASTED47:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED48:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS49:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS50:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS51:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES52:%.*]] = alloca [4 x i64], align 4
-// CHECK19-NEXT:    [[_TMP53:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_54:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_55:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[N_CASTED34:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS35:%.*]] = alloca [3 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS36:%.*]] = alloca [3 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS37:%.*]] = alloca [3 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES38:%.*]] = alloca [3 x i64], align 4
+// CHECK19-NEXT:    [[_TMP39:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_40:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_41:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_49:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[N_CASTED50:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED51:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS52:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS53:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS54:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES55:%.*]] = alloca [4 x i64], align 4
+// CHECK19-NEXT:    [[_TMP56:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_57:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_58:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK19-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK19-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -11320,262 +12170,347 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK19-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK19-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK19-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK19-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK19-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK19-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK19-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK19-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK19-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
-// CHECK19-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK19-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK19-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK19-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
-// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK19-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK19-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK19-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK19-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK19-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK19-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK19-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK19-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK19-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK19-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK19-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK19-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
+// CHECK19-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK19-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK19-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK19-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK19-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK19-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK19-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK19-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK19-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK19-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK19-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK19-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK19-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK19-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK19-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP59]])
-// CHECK19-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK19-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK19:       omp_offload.failed15:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK19:       omp_offload.cont16:
-// CHECK19-NEXT:    [[TMP62:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP62]], i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK19-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP63]], i32* [[N_CASTED18]], align 4
-// CHECK19-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N_CASTED18]], align 4
-// CHECK19-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK19-NEXT:    store i32 [[TMP65]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK19-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK19-NEXT:    [[TMP67:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP68:%.*]] = sext i32 [[TMP67]] to i64
-// CHECK19-NEXT:    [[TMP69:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP69]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
-// CHECK19-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP64]], i32* [[TMP71]], align 4
-// CHECK19-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP64]], i32* [[TMP73]], align 4
-// CHECK19-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP74]], align 4
-// CHECK19-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP76]], align 4
-// CHECK19-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP78]], align 4
-// CHECK19-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP79]], align 4
-// CHECK19-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 4
-// CHECK19-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 4
-// CHECK19-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 2
-// CHECK19-NEXT:    store i64 [[TMP68]], i64* [[TMP84]], align 4
-// CHECK19-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP85]], align 4
-// CHECK19-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK19-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK19-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP67]])
+// CHECK19-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK19-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK19-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK19-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK19-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK19-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK19-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK19-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK19-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK19-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK19-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK19:       omp_offload.failed16:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK19:       omp_offload.cont17:
+// CHECK19-NEXT:    [[TMP78:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP78]], i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK19-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP79]], i32* [[N_CASTED19]], align 4
+// CHECK19-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_CASTED19]], align 4
+// CHECK19-NEXT:    [[TMP81:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK19-NEXT:    store i32 [[TMP81]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK19-NEXT:    [[TMP82:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK19-NEXT:    [[TMP83:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP84:%.*]] = sext i32 [[TMP83]] to i64
+// CHECK19-NEXT:    [[TMP85:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES23]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP85]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
+// CHECK19-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP66]], i32* [[TMP87]], align 4
-// CHECK19-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK19-NEXT:    store i32 [[TMP80]], i32* [[TMP87]], align 4
+// CHECK19-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP66]], i32* [[TMP89]], align 4
-// CHECK19-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK19-NEXT:    store i32 [[TMP80]], i32* [[TMP89]], align 4
+// CHECK19-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
 // CHECK19-NEXT:    store i8* null, i8** [[TMP90]], align 4
-// CHECK19-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP94:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP94]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK19-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK19-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP95]], 0
-// CHECK19-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
-// CHECK19-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
-// CHECK19-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK19-NEXT:    [[TMP96:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK19-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP96]], 1
-// CHECK19-NEXT:    [[TMP97:%.*]] = zext i32 [[ADD29]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP97]])
-// CHECK19-NEXT:    [[TMP98:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, i32 4, i8** [[TMP91]], i8** [[TMP92]], i64* [[TMP93]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP99:%.*]] = icmp ne i32 [[TMP98]], 0
-// CHECK19-NEXT:    br i1 [[TMP99]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK19:       omp_offload.failed30:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i32 [[TMP64]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP66]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK19:       omp_offload.cont31:
-// CHECK19-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP100]], i32* [[N_CASTED32]], align 4
-// CHECK19-NEXT:    [[TMP101:%.*]] = load i32, i32* [[N_CASTED32]], align 4
-// CHECK19-NEXT:    [[TMP102:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP103:%.*]] = sext i32 [[TMP102]] to i64
-// CHECK19-NEXT:    [[TMP104:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES36]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP104]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
-// CHECK19-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP106:%.*]] = bitcast i8** [[TMP105]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP101]], i32* [[TMP106]], align 4
-// CHECK19-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP101]], i32* [[TMP108]], align 4
-// CHECK19-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP109]], align 4
-// CHECK19-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP111:%.*]] = bitcast i8** [[TMP110]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP111]], align 4
-// CHECK19-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP113]], align 4
-// CHECK19-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP114]], align 4
-// CHECK19-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP116]], align 4
-// CHECK19-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP118]], align 4
-// CHECK19-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES36]], i32 0, i32 2
-// CHECK19-NEXT:    store i64 [[TMP103]], i64* [[TMP119]], align 4
-// CHECK19-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP120]], align 4
-// CHECK19-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES36]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP92]], align 4
+// CHECK19-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP94]], align 4
+// CHECK19-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP95]], align 4
+// CHECK19-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 4
+// CHECK19-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 4
+// CHECK19-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 2
+// CHECK19-NEXT:    store i64 [[TMP84]], i64* [[TMP100]], align 4
+// CHECK19-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP101]], align 4
+// CHECK19-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP82]], i32* [[TMP103]], align 4
+// CHECK19-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP82]], i32* [[TMP105]], align 4
+// CHECK19-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 3
+// CHECK19-NEXT:    store i8* null, i8** [[TMP106]], align 4
+// CHECK19-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP110:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP110]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK19-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK19-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP111]], 0
+// CHECK19-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
+// CHECK19-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
+// CHECK19-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK19-NEXT:    [[TMP112:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK19-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP112]], 1
+// CHECK19-NEXT:    [[TMP113:%.*]] = zext i32 [[ADD30]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP113]])
+// CHECK19-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP114]], align 4
+// CHECK19-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 4, i32* [[TMP115]], align 4
+// CHECK19-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK19-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP108]], i8*** [[TMP117]], align 4
+// CHECK19-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP109]], i64** [[TMP118]], align 4
+// CHECK19-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP119]], align 4
+// CHECK19-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK19-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP121]], align 4
+// CHECK19-NEXT:    [[TMP122:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK19-NEXT:    [[TMP123:%.*]] = icmp ne i32 [[TMP122]], 0
+// CHECK19-NEXT:    br i1 [[TMP123]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK19:       omp_offload.failed32:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i32 [[TMP80]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP82]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK19:       omp_offload.cont33:
 // CHECK19-NEXT:    [[TMP124:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP124]], i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK19-NEXT:    [[TMP125:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK19-NEXT:    [[SUB40:%.*]] = sub nsw i32 [[TMP125]], 0
-// CHECK19-NEXT:    [[DIV41:%.*]] = sdiv i32 [[SUB40]], 1
-// CHECK19-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[DIV41]], 1
-// CHECK19-NEXT:    store i32 [[SUB42]], i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK19-NEXT:    [[TMP126:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK19-NEXT:    [[ADD43:%.*]] = add nsw i32 [[TMP126]], 1
-// CHECK19-NEXT:    [[TMP127:%.*]] = zext i32 [[ADD43]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP127]])
-// CHECK19-NEXT:    [[TMP128:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, i32 3, i8** [[TMP121]], i8** [[TMP122]], i64* [[TMP123]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP129:%.*]] = icmp ne i32 [[TMP128]], 0
-// CHECK19-NEXT:    br i1 [[TMP129]], label [[OMP_OFFLOAD_FAILED44:%.*]], label [[OMP_OFFLOAD_CONT45:%.*]]
-// CHECK19:       omp_offload.failed44:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i32 [[TMP101]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT45]]
-// CHECK19:       omp_offload.cont45:
-// CHECK19-NEXT:    [[TMP130:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP130]], i32* [[DOTCAPTURE_EXPR_46]], align 4
-// CHECK19-NEXT:    [[TMP131:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP131]], i32* [[N_CASTED47]], align 4
-// CHECK19-NEXT:    [[TMP132:%.*]] = load i32, i32* [[N_CASTED47]], align 4
-// CHECK19-NEXT:    [[TMP133:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_46]], align 4
-// CHECK19-NEXT:    store i32 [[TMP133]], i32* [[DOTCAPTURE_EXPR__CASTED48]], align 4
-// CHECK19-NEXT:    [[TMP134:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED48]], align 4
-// CHECK19-NEXT:    [[TMP135:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP136:%.*]] = sext i32 [[TMP135]] to i64
-// CHECK19-NEXT:    [[TMP137:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES52]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP137]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
-// CHECK19-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP132]], i32* [[TMP139]], align 4
-// CHECK19-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP132]], i32* [[TMP141]], align 4
-// CHECK19-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP142]], align 4
-// CHECK19-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP144]], align 4
-// CHECK19-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP146]], align 4
-// CHECK19-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP147]], align 4
-// CHECK19-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP149]], align 4
-// CHECK19-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP151]], align 4
-// CHECK19-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES52]], i32 0, i32 2
-// CHECK19-NEXT:    store i64 [[TMP136]], i64* [[TMP152]], align 4
-// CHECK19-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP153]], align 4
-// CHECK19-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP155:%.*]] = bitcast i8** [[TMP154]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP134]], i32* [[TMP155]], align 4
-// CHECK19-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP134]], i32* [[TMP157]], align 4
-// CHECK19-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 3
-// CHECK19-NEXT:    store i8* null, i8** [[TMP158]], align 4
-// CHECK19-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES52]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP162:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP162]], i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK19-NEXT:    [[TMP163:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK19-NEXT:    [[SUB56:%.*]] = sub nsw i32 [[TMP163]], 0
-// CHECK19-NEXT:    [[DIV57:%.*]] = sdiv i32 [[SUB56]], 1
-// CHECK19-NEXT:    [[SUB58:%.*]] = sub nsw i32 [[DIV57]], 1
-// CHECK19-NEXT:    store i32 [[SUB58]], i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK19-NEXT:    [[TMP164:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK19-NEXT:    [[ADD59:%.*]] = add nsw i32 [[TMP164]], 1
-// CHECK19-NEXT:    [[TMP165:%.*]] = zext i32 [[ADD59]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP165]])
-// CHECK19-NEXT:    [[TMP166:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, i32 4, i8** [[TMP159]], i8** [[TMP160]], i64* [[TMP161]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP167:%.*]] = icmp ne i32 [[TMP166]], 0
-// CHECK19-NEXT:    br i1 [[TMP167]], label [[OMP_OFFLOAD_FAILED60:%.*]], label [[OMP_OFFLOAD_CONT61:%.*]]
-// CHECK19:       omp_offload.failed60:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i32 [[TMP132]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP134]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT61]]
-// CHECK19:       omp_offload.cont61:
-// CHECK19-NEXT:    [[TMP168:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK19-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP168]])
+// CHECK19-NEXT:    store i32 [[TMP124]], i32* [[N_CASTED34]], align 4
+// CHECK19-NEXT:    [[TMP125:%.*]] = load i32, i32* [[N_CASTED34]], align 4
+// CHECK19-NEXT:    [[TMP126:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP127:%.*]] = sext i32 [[TMP126]] to i64
+// CHECK19-NEXT:    [[TMP128:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES38]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP128]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
+// CHECK19-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP125]], i32* [[TMP130]], align 4
+// CHECK19-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP125]], i32* [[TMP132]], align 4
+// CHECK19-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS37]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP133]], align 4
+// CHECK19-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP135]], align 4
+// CHECK19-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP137]], align 4
+// CHECK19-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS37]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP138]], align 4
+// CHECK19-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP140]], align 4
+// CHECK19-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP142]], align 4
+// CHECK19-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES38]], i32 0, i32 2
+// CHECK19-NEXT:    store i64 [[TMP127]], i64* [[TMP143]], align 4
+// CHECK19-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS37]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP144]], align 4
+// CHECK19-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES38]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP148:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP148]], i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK19-NEXT:    [[TMP149:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK19-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[TMP149]], 0
+// CHECK19-NEXT:    [[DIV43:%.*]] = sdiv i32 [[SUB42]], 1
+// CHECK19-NEXT:    [[SUB44:%.*]] = sub nsw i32 [[DIV43]], 1
+// CHECK19-NEXT:    store i32 [[SUB44]], i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK19-NEXT:    [[TMP150:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK19-NEXT:    [[ADD45:%.*]] = add nsw i32 [[TMP150]], 1
+// CHECK19-NEXT:    [[TMP151:%.*]] = zext i32 [[ADD45]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP151]])
+// CHECK19-NEXT:    [[KERNEL_ARGS46:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP152]], align 4
+// CHECK19-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP153]], align 4
+// CHECK19-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP145]], i8*** [[TMP154]], align 4
+// CHECK19-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP146]], i8*** [[TMP155]], align 4
+// CHECK19-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP147]], i64** [[TMP156]], align 4
+// CHECK19-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP157]], align 4
+// CHECK19-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP158]], align 4
+// CHECK19-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP159]], align 4
+// CHECK19-NEXT:    [[TMP160:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]])
+// CHECK19-NEXT:    [[TMP161:%.*]] = icmp ne i32 [[TMP160]], 0
+// CHECK19-NEXT:    br i1 [[TMP161]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
+// CHECK19:       omp_offload.failed47:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i32 [[TMP125]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
+// CHECK19:       omp_offload.cont48:
+// CHECK19-NEXT:    [[TMP162:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP162]], i32* [[DOTCAPTURE_EXPR_49]], align 4
+// CHECK19-NEXT:    [[TMP163:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP163]], i32* [[N_CASTED50]], align 4
+// CHECK19-NEXT:    [[TMP164:%.*]] = load i32, i32* [[N_CASTED50]], align 4
+// CHECK19-NEXT:    [[TMP165:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_49]], align 4
+// CHECK19-NEXT:    store i32 [[TMP165]], i32* [[DOTCAPTURE_EXPR__CASTED51]], align 4
+// CHECK19-NEXT:    [[TMP166:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED51]], align 4
+// CHECK19-NEXT:    [[TMP167:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP168:%.*]] = sext i32 [[TMP167]] to i64
+// CHECK19-NEXT:    [[TMP169:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES55]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP169]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
+// CHECK19-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP171:%.*]] = bitcast i8** [[TMP170]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP164]], i32* [[TMP171]], align 4
+// CHECK19-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP173:%.*]] = bitcast i8** [[TMP172]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP164]], i32* [[TMP173]], align 4
+// CHECK19-NEXT:    [[TMP174:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP174]], align 4
+// CHECK19-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP176:%.*]] = bitcast i8** [[TMP175]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP176]], align 4
+// CHECK19-NEXT:    [[TMP177:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP178:%.*]] = bitcast i8** [[TMP177]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP178]], align 4
+// CHECK19-NEXT:    [[TMP179:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP179]], align 4
+// CHECK19-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP181]], align 4
+// CHECK19-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP183:%.*]] = bitcast i8** [[TMP182]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP183]], align 4
+// CHECK19-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES55]], i32 0, i32 2
+// CHECK19-NEXT:    store i64 [[TMP168]], i64* [[TMP184]], align 4
+// CHECK19-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP185]], align 4
+// CHECK19-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP187:%.*]] = bitcast i8** [[TMP186]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP166]], i32* [[TMP187]], align 4
+// CHECK19-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP189:%.*]] = bitcast i8** [[TMP188]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP166]], i32* [[TMP189]], align 4
+// CHECK19-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 3
+// CHECK19-NEXT:    store i8* null, i8** [[TMP190]], align 4
+// CHECK19-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES55]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP194:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP194]], i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK19-NEXT:    [[TMP195:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK19-NEXT:    [[SUB59:%.*]] = sub nsw i32 [[TMP195]], 0
+// CHECK19-NEXT:    [[DIV60:%.*]] = sdiv i32 [[SUB59]], 1
+// CHECK19-NEXT:    [[SUB61:%.*]] = sub nsw i32 [[DIV60]], 1
+// CHECK19-NEXT:    store i32 [[SUB61]], i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK19-NEXT:    [[TMP196:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK19-NEXT:    [[ADD62:%.*]] = add nsw i32 [[TMP196]], 1
+// CHECK19-NEXT:    [[TMP197:%.*]] = zext i32 [[ADD62]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP197]])
+// CHECK19-NEXT:    [[KERNEL_ARGS63:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP198]], align 4
+// CHECK19-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 4, i32* [[TMP199]], align 4
+// CHECK19-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP191]], i8*** [[TMP200]], align 4
+// CHECK19-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP192]], i8*** [[TMP201]], align 4
+// CHECK19-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP193]], i64** [[TMP202]], align 4
+// CHECK19-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP203]], align 4
+// CHECK19-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP204]], align 4
+// CHECK19-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP205]], align 4
+// CHECK19-NEXT:    [[TMP206:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]])
+// CHECK19-NEXT:    [[TMP207:%.*]] = icmp ne i32 [[TMP206]], 0
+// CHECK19-NEXT:    br i1 [[TMP207]], label [[OMP_OFFLOAD_FAILED64:%.*]], label [[OMP_OFFLOAD_CONT65:%.*]]
+// CHECK19:       omp_offload.failed64:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i32 [[TMP164]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP166]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT65]]
+// CHECK19:       omp_offload.cont65:
+// CHECK19-NEXT:    [[TMP208:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK19-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP208]])
 // CHECK19-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK19-NEXT:    [[TMP169:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP169]])
-// CHECK19-NEXT:    [[TMP170:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK19-NEXT:    ret i32 [[TMP170]]
+// CHECK19-NEXT:    [[TMP209:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP209]])
+// CHECK19-NEXT:    [[TMP210:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK19-NEXT:    ret i32 [[TMP210]]
 //
 //
 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139
@@ -12700,20 +13635,20 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 4
-// CHECK19-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED20:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 4
+// CHECK19-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED23:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK19-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK19-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -12727,111 +13662,196 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK19-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK19-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK19-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK19-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK19-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK19-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK19-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK19-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
-// CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK19-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK19-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK19-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK19-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK19:       omp_offload.failed5:
+// CHECK19-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK19-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK19-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK19-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK19-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK19:       omp_offload.failed6:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK19:       omp_offload.cont6:
-// CHECK19-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP18]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK19-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK19-NEXT:    store i32 [[TMP19]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK19-NEXT:    [[TMP20:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP22]], align 4
-// CHECK19-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP24]], align 4
-// CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP25]], align 4
-// CHECK19-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP20]], i32* [[TMP27]], align 4
-// CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP20]], i32* [[TMP29]], align 4
-// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP30]], align 4
-// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, i32 2, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK19-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK19:       omp_offload.failed11:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i32 [[TMP20]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK19:       omp_offload.cont12:
-// CHECK19-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP36]], align 4
-// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK19:       omp_offload.cont7:
+// CHECK19-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP34]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK19-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK19-NEXT:    store i32 [[TMP35]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK19-NEXT:    [[TMP36:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [10 x i32]**
 // CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP38]], align 4
-// CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP39]], align 4
-// CHECK19-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP40]], align 4
+// CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP41]], align 4
+// CHECK19-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP36]], i32* [[TMP43]], align 4
+// CHECK19-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP36]], i32* [[TMP45]], align 4
+// CHECK19-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP46]], align 4
+// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
+// CHECK19-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP49]], align 4
+// CHECK19-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 2, i32* [[TMP50]], align 4
+// CHECK19-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 4
+// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP48]], i8*** [[TMP52]], align 4
+// CHECK19-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP53]], align 4
+// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP54]], align 4
+// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP55]], align 4
+// CHECK19-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP56]], align 4
+// CHECK19-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK19-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
+// CHECK19-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK19:       omp_offload.failed13:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i32 [[TMP36]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK19:       omp_offload.cont14:
+// CHECK19-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP60]], align 4
+// CHECK19-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP62]], align 4
+// CHECK19-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP63]], align 4
+// CHECK19-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, i32 1, i8** [[TMP40]], i8** [[TMP41]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK19-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK19:       omp_offload.failed17:
+// CHECK19-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK19-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK19-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK19-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 4
+// CHECK19-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK19-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP71]], align 4
+// CHECK19-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK19-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP73]], align 4
+// CHECK19-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK19-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK19-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK19:       omp_offload.failed20:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK19:       omp_offload.cont18:
-// CHECK19-NEXT:    [[TMP44:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP44]], i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK19-NEXT:    [[TMP45:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK19-NEXT:    store i32 [[TMP45]], i32* [[DOTCAPTURE_EXPR__CASTED20]], align 4
-// CHECK19-NEXT:    [[TMP46:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED20]], align 4
-// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP48]], align 4
-// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP50]], align 4
-// CHECK19-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP51]], align 4
-// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP46]], i32* [[TMP53]], align 4
-// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP46]], i32* [[TMP55]], align 4
-// CHECK19-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP56]], align 4
-// CHECK19-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK19:       omp_offload.cont21:
+// CHECK19-NEXT:    [[TMP76:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP76]], i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK19-NEXT:    [[TMP77:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK19-NEXT:    store i32 [[TMP77]], i32* [[DOTCAPTURE_EXPR__CASTED23]], align 4
+// CHECK19-NEXT:    [[TMP78:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED23]], align 4
+// CHECK19-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP80]], align 4
+// CHECK19-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP82]], align 4
+// CHECK19-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP83]], align 4
+// CHECK19-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP78]], i32* [[TMP85]], align 4
+// CHECK19-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP78]], i32* [[TMP87]], align 4
+// CHECK19-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP88]], align 4
+// CHECK19-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, i32 2, i8** [[TMP57]], i8** [[TMP58]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK19-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
-// CHECK19:       omp_offload.failed25:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i32 [[TMP46]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
-// CHECK19:       omp_offload.cont26:
+// CHECK19-NEXT:    [[KERNEL_ARGS28:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP91]], align 4
+// CHECK19-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 2, i32* [[TMP92]], align 4
+// CHECK19-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP89]], i8*** [[TMP93]], align 4
+// CHECK19-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP90]], i8*** [[TMP94]], align 4
+// CHECK19-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP95]], align 4
+// CHECK19-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP96]], align 4
+// CHECK19-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP97]], align 4
+// CHECK19-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP98]], align 4
+// CHECK19-NEXT:    [[TMP99:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]])
+// CHECK19-NEXT:    [[TMP100:%.*]] = icmp ne i32 [[TMP99]], 0
+// CHECK19-NEXT:    br i1 [[TMP100]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
+// CHECK19:       omp_offload.failed29:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i32 [[TMP78]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
+// CHECK19:       omp_offload.cont30:
 // CHECK19-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_codegen.cpp
index ba25d49..9807ffc 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_codegen.cpp
@@ -179,65 +179,99 @@ int target_teams_fun(int *g){
 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP42]], 1
 // CHECK1-NEXT:    [[TMP43:%.*]] = zext i32 [[ADD]] to i64
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 [[TMP43]])
-// CHECK1-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l50.region_id, i32 5, i8** [[TMP36]], i8** [[TMP37]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP38]], i32 [[TMP39]])
-// CHECK1-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
-// CHECK1-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP44]], align 4
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP45]], align 4
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP36]], i8*** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP37]], i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP49]], align 8
+// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP50]], align 8
+// CHECK1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP51]], align 8
+// CHECK1-NEXT:    [[TMP52:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 [[TMP38]], i32 [[TMP39]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l50.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP53:%.*]] = icmp ne i32 [[TMP52]], 0
+// CHECK1-NEXT:    br i1 [[TMP53]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l50(i64 [[TMP4]], i64 [[TMP6]], [1000 x i32]* [[A]], i64 [[TMP8]], i64 [[TMP10]]) #[[ATTR4:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP46:%.*]] = load i32, i32* [[N]], align 4
+// CHECK1-NEXT:    [[TMP54:%.*]] = load i32, i32* [[N]], align 4
 // CHECK1-NEXT:    [[CONV11:%.*]] = bitcast i64* [[N_CASTED10]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP46]], i32* [[CONV11]], align 4
-// CHECK1-NEXT:    [[TMP47:%.*]] = load i64, i64* [[N_CASTED10]], align 8
-// CHECK1-NEXT:    [[TMP48:%.*]] = load i32*, i32** [[G_ADDR]], align 8
-// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP47]], i64* [[TMP50]], align 8
-// CHECK1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP47]], i64* [[TMP52]], align 8
-// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP53]], align 8
-// CHECK1-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to [1000 x i32]**
-// CHECK1-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP55]], align 8
-// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to [1000 x i32]**
-// CHECK1-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP57]], align 8
-// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP58]], align 8
-// CHECK1-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to i32**
-// CHECK1-NEXT:    store i32* [[TMP48]], i32** [[TMP60]], align 8
-// CHECK1-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i32**
-// CHECK1-NEXT:    store i32* [[TMP48]], i32** [[TMP62]], align 8
-// CHECK1-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP63]], align 8
-// CHECK1-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP66:%.*]] = load i32, i32* [[N]], align 4
-// CHECK1-NEXT:    store i32 [[TMP66]], i32* [[DOTCAPTURE_EXPR_16]], align 4
-// CHECK1-NEXT:    [[TMP67:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_16]], align 4
-// CHECK1-NEXT:    [[SUB18:%.*]] = sub nsw i32 [[TMP67]], 0
+// CHECK1-NEXT:    store i32 [[TMP54]], i32* [[CONV11]], align 4
+// CHECK1-NEXT:    [[TMP55:%.*]] = load i64, i64* [[N_CASTED10]], align 8
+// CHECK1-NEXT:    [[TMP56:%.*]] = load i32*, i32** [[G_ADDR]], align 8
+// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP55]], i64* [[TMP58]], align 8
+// CHECK1-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP55]], i64* [[TMP60]], align 8
+// CHECK1-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP61]], align 8
+// CHECK1-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP63:%.*]] = bitcast i8** [[TMP62]] to [1000 x i32]**
+// CHECK1-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP63]], align 8
+// CHECK1-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP65:%.*]] = bitcast i8** [[TMP64]] to [1000 x i32]**
+// CHECK1-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP65]], align 8
+// CHECK1-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP66]], align 8
+// CHECK1-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i32**
+// CHECK1-NEXT:    store i32* [[TMP56]], i32** [[TMP68]], align 8
+// CHECK1-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32**
+// CHECK1-NEXT:    store i32* [[TMP56]], i32** [[TMP70]], align 8
+// CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i64 0, i64 2
+// CHECK1-NEXT:    store i8* null, i8** [[TMP71]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP74:%.*]] = load i32, i32* [[N]], align 4
+// CHECK1-NEXT:    store i32 [[TMP74]], i32* [[DOTCAPTURE_EXPR_16]], align 4
+// CHECK1-NEXT:    [[TMP75:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_16]], align 4
+// CHECK1-NEXT:    [[SUB18:%.*]] = sub nsw i32 [[TMP75]], 0
 // CHECK1-NEXT:    [[DIV19:%.*]] = sdiv i32 [[SUB18]], 1
 // CHECK1-NEXT:    [[SUB20:%.*]] = sub nsw i32 [[DIV19]], 1
 // CHECK1-NEXT:    store i32 [[SUB20]], i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK1-NEXT:    [[TMP68:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK1-NEXT:    [[ADD21:%.*]] = add nsw i32 [[TMP68]], 1
-// CHECK1-NEXT:    [[TMP69:%.*]] = zext i32 [[ADD21]] to i64
-// CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 [[TMP69]])
-// CHECK1-NEXT:    [[TMP70:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l56.region_id, i32 3, i8** [[TMP64]], i8** [[TMP65]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP71:%.*]] = icmp ne i32 [[TMP70]], 0
-// CHECK1-NEXT:    br i1 [[TMP71]], label [[OMP_OFFLOAD_FAILED22:%.*]], label [[OMP_OFFLOAD_CONT23:%.*]]
-// CHECK1:       omp_offload.failed22:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l56(i64 [[TMP47]], [1000 x i32]* [[A]], i32* [[TMP48]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT23]]
-// CHECK1:       omp_offload.cont23:
+// CHECK1-NEXT:    [[TMP76:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_17]], align 4
+// CHECK1-NEXT:    [[ADD21:%.*]] = add nsw i32 [[TMP76]], 1
+// CHECK1-NEXT:    [[TMP77:%.*]] = zext i32 [[ADD21]] to i64
+// CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 [[TMP77]])
+// CHECK1-NEXT:    [[KERNEL_ARGS22:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS22]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP78]], align 4
+// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS22]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP79]], align 4
+// CHECK1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS22]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP72]], i8*** [[TMP80]], align 8
+// CHECK1-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS22]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP73]], i8*** [[TMP81]], align 8
+// CHECK1-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS22]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP82]], align 8
+// CHECK1-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS22]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP83]], align 8
+// CHECK1-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS22]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP84]], align 8
+// CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS22]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP85]], align 8
+// CHECK1-NEXT:    [[TMP86:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS22]])
+// CHECK1-NEXT:    [[TMP87:%.*]] = icmp ne i32 [[TMP86]], 0
+// CHECK1-NEXT:    br i1 [[TMP87]], label [[OMP_OFFLOAD_FAILED23:%.*]], label [[OMP_OFFLOAD_CONT24:%.*]]
+// CHECK1:       omp_offload.failed23:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l56(i64 [[TMP55]], [1000 x i32]* [[A]], i32* [[TMP56]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT24]]
+// CHECK1:       omp_offload.cont24:
 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [1000 x i32], [1000 x i32]* [[A]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP72:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP72]]
+// CHECK1-NEXT:    [[TMP88:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP88]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l50
@@ -886,64 +920,98 @@ int target_teams_fun(int *g){
 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP42]], 1
 // CHECK3-NEXT:    [[TMP43:%.*]] = zext i32 [[ADD]] to i64
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 [[TMP43]])
-// CHECK3-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l50.region_id, i32 5, i8** [[TMP36]], i8** [[TMP37]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP38]], i32 [[TMP39]])
-// CHECK3-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
-// CHECK3-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP36]], i8*** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP37]], i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP49]], align 4
+// CHECK3-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP50]], align 4
+// CHECK3-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP51]], align 4
+// CHECK3-NEXT:    [[TMP52:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 [[TMP38]], i32 [[TMP39]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l50.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP53:%.*]] = icmp ne i32 [[TMP52]], 0
+// CHECK3-NEXT:    br i1 [[TMP53]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l50(i32 [[TMP4]], i32 [[TMP6]], [1000 x i32]* [[A]], i32 [[TMP8]], i32 [[TMP10]]) #[[ATTR4:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP46:%.*]] = load i32, i32* [[N]], align 4
-// CHECK3-NEXT:    store i32 [[TMP46]], i32* [[N_CASTED7]], align 4
-// CHECK3-NEXT:    [[TMP47:%.*]] = load i32, i32* [[N_CASTED7]], align 4
-// CHECK3-NEXT:    [[TMP48:%.*]] = load i32*, i32** [[G_ADDR]], align 4
-// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP47]], i32* [[TMP50]], align 4
-// CHECK3-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP47]], i32* [[TMP52]], align 4
-// CHECK3-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP53]], align 4
-// CHECK3-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to [1000 x i32]**
-// CHECK3-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP55]], align 4
-// CHECK3-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to [1000 x i32]**
-// CHECK3-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP57]], align 4
-// CHECK3-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP58]], align 4
-// CHECK3-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to i32**
-// CHECK3-NEXT:    store i32* [[TMP48]], i32** [[TMP60]], align 4
-// CHECK3-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to i32**
-// CHECK3-NEXT:    store i32* [[TMP48]], i32** [[TMP62]], align 4
-// CHECK3-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP63]], align 4
-// CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP66:%.*]] = load i32, i32* [[N]], align 4
-// CHECK3-NEXT:    store i32 [[TMP66]], i32* [[DOTCAPTURE_EXPR_12]], align 4
-// CHECK3-NEXT:    [[TMP67:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_12]], align 4
-// CHECK3-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[TMP67]], 0
+// CHECK3-NEXT:    [[TMP54:%.*]] = load i32, i32* [[N]], align 4
+// CHECK3-NEXT:    store i32 [[TMP54]], i32* [[N_CASTED7]], align 4
+// CHECK3-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N_CASTED7]], align 4
+// CHECK3-NEXT:    [[TMP56:%.*]] = load i32*, i32** [[G_ADDR]], align 4
+// CHECK3-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP55]], i32* [[TMP58]], align 4
+// CHECK3-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP55]], i32* [[TMP60]], align 4
+// CHECK3-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP61]], align 4
+// CHECK3-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP63:%.*]] = bitcast i8** [[TMP62]] to [1000 x i32]**
+// CHECK3-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP63]], align 4
+// CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP65:%.*]] = bitcast i8** [[TMP64]] to [1000 x i32]**
+// CHECK3-NEXT:    store [1000 x i32]* [[A]], [1000 x i32]** [[TMP65]], align 4
+// CHECK3-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP66]], align 4
+// CHECK3-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i32**
+// CHECK3-NEXT:    store i32* [[TMP56]], i32** [[TMP68]], align 4
+// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32**
+// CHECK3-NEXT:    store i32* [[TMP56]], i32** [[TMP70]], align 4
+// CHECK3-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 2
+// CHECK3-NEXT:    store i8* null, i8** [[TMP71]], align 4
+// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP74:%.*]] = load i32, i32* [[N]], align 4
+// CHECK3-NEXT:    store i32 [[TMP74]], i32* [[DOTCAPTURE_EXPR_12]], align 4
+// CHECK3-NEXT:    [[TMP75:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_12]], align 4
+// CHECK3-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[TMP75]], 0
 // CHECK3-NEXT:    [[DIV15:%.*]] = sdiv i32 [[SUB14]], 1
 // CHECK3-NEXT:    [[SUB16:%.*]] = sub nsw i32 [[DIV15]], 1
 // CHECK3-NEXT:    store i32 [[SUB16]], i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK3-NEXT:    [[TMP68:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK3-NEXT:    [[ADD17:%.*]] = add nsw i32 [[TMP68]], 1
-// CHECK3-NEXT:    [[TMP69:%.*]] = zext i32 [[ADD17]] to i64
-// CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 [[TMP69]])
-// CHECK3-NEXT:    [[TMP70:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l56.region_id, i32 3, i8** [[TMP64]], i8** [[TMP65]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP71:%.*]] = icmp ne i32 [[TMP70]], 0
-// CHECK3-NEXT:    br i1 [[TMP71]], label [[OMP_OFFLOAD_FAILED18:%.*]], label [[OMP_OFFLOAD_CONT19:%.*]]
-// CHECK3:       omp_offload.failed18:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l56(i32 [[TMP47]], [1000 x i32]* [[A]], i32* [[TMP48]]) #[[ATTR4]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT19]]
-// CHECK3:       omp_offload.cont19:
+// CHECK3-NEXT:    [[TMP76:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
+// CHECK3-NEXT:    [[ADD17:%.*]] = add nsw i32 [[TMP76]], 1
+// CHECK3-NEXT:    [[TMP77:%.*]] = zext i32 [[ADD17]] to i64
+// CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 [[TMP77]])
+// CHECK3-NEXT:    [[KERNEL_ARGS18:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP78]], align 4
+// CHECK3-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP79]], align 4
+// CHECK3-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP72]], i8*** [[TMP80]], align 4
+// CHECK3-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP73]], i8*** [[TMP81]], align 4
+// CHECK3-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP82]], align 4
+// CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP83]], align 4
+// CHECK3-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP84]], align 4
+// CHECK3-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP85]], align 4
+// CHECK3-NEXT:    [[TMP86:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]])
+// CHECK3-NEXT:    [[TMP87:%.*]] = icmp ne i32 [[TMP86]], 0
+// CHECK3-NEXT:    br i1 [[TMP87]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
+// CHECK3:       omp_offload.failed19:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l56(i32 [[TMP55]], [1000 x i32]* [[A]], i32* [[TMP56]]) #[[ATTR4]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
+// CHECK3:       omp_offload.cont20:
 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [1000 x i32], [1000 x i32]* [[A]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP72:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP72]]
+// CHECK3-NEXT:    [[TMP88:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP88]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16target_teams_funPi_l50

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_collapse_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_collapse_codegen.cpp
index 1aae008..32b1b9e 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_collapse_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_collapse_codegen.cpp
@@ -130,9 +130,26 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 56088)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -140,8 +157,8 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i64 0, i64 0
 // CHECK1-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP9]]
+// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -365,9 +382,26 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 56088)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -375,8 +409,8 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP9]]
+// CHECK3-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -815,20 +849,37 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP46:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_4]], align 8
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP46]], 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK9-NEXT:    [[TMP47:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, i32 5, i8** [[TMP39]], i8** [[TMP40]], i64* [[TMP41]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP48:%.*]] = icmp ne i32 [[TMP47]], 0
-// CHECK9-NEXT:    br i1 [[TMP48]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP47]], align 4
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP48]], align 4
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 8
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP40]], i8*** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP41]], i64** [[TMP51]], align 8
+// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP56:%.*]] = icmp ne i32 [[TMP55]], 0
+// CHECK9-NEXT:    br i1 [[TMP56]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81(i64 [[TMP7]], i64 [[TMP9]], i64 [[TMP1]], i64 [[TMP3]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP49:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP49]])
+// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP57]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP50:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP50]])
-// CHECK9-NEXT:    [[TMP51:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP51]]
+// CHECK9-NEXT:    [[TMP58:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP58]])
+// CHECK9-NEXT:    [[TMP59:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP59]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81
@@ -1199,9 +1250,26 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 20)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1507,20 +1575,37 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP45:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_3]], align 8
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP45]], 1
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK11-NEXT:    [[TMP46:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, i32 5, i8** [[TMP38]], i8** [[TMP39]], i64* [[TMP40]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP47:%.*]] = icmp ne i32 [[TMP46]], 0
-// CHECK11-NEXT:    br i1 [[TMP47]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP38]], i8*** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP40]], i64** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP55:%.*]] = icmp ne i32 [[TMP54]], 0
+// CHECK11-NEXT:    br i1 [[TMP55]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81(i32 [[TMP5]], i32 [[TMP7]], i32 [[TMP0]], i32 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP48:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP48]])
+// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP56]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP49]])
-// CHECK11-NEXT:    [[TMP50:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP50]]
+// CHECK11-NEXT:    [[TMP57:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP57]])
+// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP58]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81
@@ -1883,9 +1968,26 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 20)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_depend_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_depend_codegen.cpp
index 0429d08..92668b1 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_depend_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_depend_codegen.cpp
@@ -182,7 +182,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 0
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 0, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // CHECK:       [[FAIL]]
@@ -200,7 +200,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 2
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 2, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SZT]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT]], i32 0, i32 0), i8** null, i8** null, i32 0, i32 0, i32 0, i8* null, i32 0, i8* null)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 0, i32 0, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 
 // CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_dist_schedule_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_dist_schedule_codegen.cpp
index 89aa3ee..04d884e 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_dist_schedule_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_dist_schedule_codegen.cpp
@@ -154,10 +154,10 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -172,55 +172,106 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK1:       omp_offload.failed16:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK1:       omp_offload.cont15:
-// CHECK1-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP27]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK1:       omp_offload.cont17:
+// CHECK1-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i64 0, i64 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -765,10 +816,10 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -783,55 +834,106 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK3:       omp_offload.failed8:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK3:       omp_offload.cont8:
-// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK3:       omp_offload.cont9:
+// CHECK3-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK3-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK3:       omp_offload.failed14:
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK3:       omp_offload.cont15:
-// CHECK3-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP27]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -1614,16 +1716,16 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[_TMP9:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
-// CHECK9-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES26:%.*]] = alloca [4 x i64], align 8
+// CHECK9-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_29:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -1682,142 +1784,193 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK9-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK9-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK9-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
-// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK9-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK9-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK9-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK9-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK9-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK9-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK9-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP58]])
-// CHECK9-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK9-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK9:       omp_offload.failed16:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK9:       omp_offload.cont17:
-// CHECK9-NEXT:    [[TMP61:%.*]] = load i32, i32* [[M]], align 4
-// CHECK9-NEXT:    store i32 [[TMP61]], i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK9-NEXT:    [[TMP62:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP62]], i32* [[CONV20]], align 4
-// CHECK9-NEXT:    [[TMP63:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK9-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK9-NEXT:    [[CONV21:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP64]], i32* [[CONV21]], align 4
-// CHECK9-NEXT:    [[TMP65:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK9-NEXT:    [[TMP66:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP67:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP67]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
-// CHECK9-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP63]], i64* [[TMP69]], align 8
-// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP63]], i64* [[TMP71]], align 8
-// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP72]], align 8
-// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP74]], align 8
-// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP76]], align 8
-// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP77]], align 8
-// CHECK9-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP79]], align 8
-// CHECK9-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 8
-// CHECK9-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP66]], i64* [[TMP82]], align 8
-// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP83]], align 8
-// CHECK9-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK9-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP66]])
+// CHECK9-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK9-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK9-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK9-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK9-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK9-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK9-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK9:       omp_offload.failed17:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK9:       omp_offload.cont18:
+// CHECK9-NEXT:    [[TMP77:%.*]] = load i32, i32* [[M]], align 4
+// CHECK9-NEXT:    store i32 [[TMP77]], i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK9-NEXT:    [[TMP78:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP78]], i32* [[CONV21]], align 4
+// CHECK9-NEXT:    [[TMP79:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK9-NEXT:    [[TMP80:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK9-NEXT:    [[CONV22:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP80]], i32* [[CONV22]], align 4
+// CHECK9-NEXT:    [[TMP81:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK9-NEXT:    [[TMP82:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP83:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES26]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP83]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
+// CHECK9-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP65]], i64* [[TMP85]], align 8
-// CHECK9-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK9-NEXT:    store i64 [[TMP79]], i64* [[TMP85]], align 8
+// CHECK9-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP65]], i64* [[TMP87]], align 8
-// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK9-NEXT:    store i64 [[TMP79]], i64* [[TMP87]], align 8
+// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 0
 // CHECK9-NEXT:    store i8* null, i8** [[TMP88]], align 8
-// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP92:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP92]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK9-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK9-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP93]], 0
-// CHECK9-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
-// CHECK9-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
-// CHECK9-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK9-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK9-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP94]], 1
-// CHECK9-NEXT:    [[TMP95:%.*]] = zext i32 [[ADD32]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP95]])
-// CHECK9-NEXT:    [[TMP96:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108.region_id, i32 4, i8** [[TMP89]], i8** [[TMP90]], i64* [[TMP91]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP97:%.*]] = icmp ne i32 [[TMP96]], 0
-// CHECK9-NEXT:    br i1 [[TMP97]], label [[OMP_OFFLOAD_FAILED33:%.*]], label [[OMP_OFFLOAD_CONT34:%.*]]
-// CHECK9:       omp_offload.failed33:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108(i64 [[TMP63]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP65]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT34]]
-// CHECK9:       omp_offload.cont34:
-// CHECK9-NEXT:    [[TMP98:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP98]])
+// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP90]], align 8
+// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP92]], align 8
+// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP93]], align 8
+// CHECK9-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP95]], align 8
+// CHECK9-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 8
+// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP82]], i64* [[TMP98]], align 8
+// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP99]], align 8
+// CHECK9-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP81]], i64* [[TMP101]], align 8
+// CHECK9-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP81]], i64* [[TMP103]], align 8
+// CHECK9-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP104]], align 8
+// CHECK9-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP108:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP108]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK9-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK9-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[TMP109]], 0
+// CHECK9-NEXT:    [[DIV31:%.*]] = sdiv i32 [[SUB30]], 1
+// CHECK9-NEXT:    [[SUB32:%.*]] = sub nsw i32 [[DIV31]], 1
+// CHECK9-NEXT:    store i32 [[SUB32]], i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK9-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK9-NEXT:    [[ADD33:%.*]] = add nsw i32 [[TMP110]], 1
+// CHECK9-NEXT:    [[TMP111:%.*]] = zext i32 [[ADD33]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP111]])
+// CHECK9-NEXT:    [[KERNEL_ARGS34:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP112]], align 4
+// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP113]], align 4
+// CHECK9-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 8
+// CHECK9-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP107]], i64** [[TMP116]], align 8
+// CHECK9-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP117]], align 8
+// CHECK9-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK9-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP119]], align 8
+// CHECK9-NEXT:    [[TMP120:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]])
+// CHECK9-NEXT:    [[TMP121:%.*]] = icmp ne i32 [[TMP120]], 0
+// CHECK9-NEXT:    br i1 [[TMP121]], label [[OMP_OFFLOAD_FAILED35:%.*]], label [[OMP_OFFLOAD_CONT36:%.*]]
+// CHECK9:       omp_offload.failed35:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108(i64 [[TMP79]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP81]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT36]]
+// CHECK9:       omp_offload.cont36:
+// CHECK9-NEXT:    [[TMP122:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP122]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP99:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP99]])
-// CHECK9-NEXT:    [[TMP100:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP100]]
+// CHECK9-NEXT:    [[TMP123:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP123]])
+// CHECK9-NEXT:    [[TMP124:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP124]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100
@@ -2618,10 +2771,10 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -2635,63 +2788,114 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l81.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l81([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l85.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK9-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK9:       omp_offload.failed5:
+// CHECK9-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l85.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK9-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK9-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK9:       omp_offload.failed6:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l85([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK9:       omp_offload.cont6:
-// CHECK9-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK9-NEXT:    store i32 [[TMP18]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK9-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK9:       omp_offload.cont7:
+// CHECK9-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK9-NEXT:    store i32 [[TMP34]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK9-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP19]], i32* [[CONV]], align 4
-// CHECK9-NEXT:    [[TMP20:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP22]], align 8
-// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP24]], align 8
-// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP25]], align 8
-// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP20]], i64* [[TMP27]], align 8
-// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP20]], i64* [[TMP29]], align 8
-// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP30]], align 8
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 [[TMP35]], i32* [[CONV]], align 4
+// CHECK9-NEXT:    [[TMP36:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP36]], i64* [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP36]], i64* [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP46]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, i32 2, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK9-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK9:       omp_offload.failed11:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]], i64 [[TMP20]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK9:       omp_offload.cont12:
+// CHECK9-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP49]], align 4
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 2, i32* [[TMP50]], align 4
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 8
+// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP48]], i8*** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP56]], align 8
+// CHECK9-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK9-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
+// CHECK9-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK9:       omp_offload.failed13:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]], i64 [[TMP36]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK9:       omp_offload.cont14:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -3260,16 +3464,16 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[_TMP8:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_17:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED19:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
-// CHECK11-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES23:%.*]] = alloca [4 x i64], align 4
+// CHECK11-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -3327,141 +3531,192 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK11-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK11-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK11-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK11-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK11-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
-// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK11-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK11-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK11-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK11-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK11-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK11-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP59]])
-// CHECK11-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK11-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK11:       omp_offload.failed15:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK11:       omp_offload.cont16:
-// CHECK11-NEXT:    [[TMP62:%.*]] = load i32, i32* [[M]], align 4
-// CHECK11-NEXT:    store i32 [[TMP62]], i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK11-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP63]], i32* [[N_CASTED18]], align 4
-// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N_CASTED18]], align 4
-// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK11-NEXT:    store i32 [[TMP65]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK11-NEXT:    [[TMP67:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP68:%.*]] = sext i32 [[TMP67]] to i64
-// CHECK11-NEXT:    [[TMP69:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP69]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
-// CHECK11-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[TMP71]], align 4
-// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[TMP73]], align 4
-// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP74]], align 4
-// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP76]], align 4
-// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP78]], align 4
-// CHECK11-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP79]], align 4
-// CHECK11-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 4
-// CHECK11-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 4
-// CHECK11-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP68]], i64* [[TMP84]], align 4
-// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP85]], align 4
-// CHECK11-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK11-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP67]])
+// CHECK11-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK11-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK11-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK11-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK11-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK11-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK11-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK11:       omp_offload.failed16:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l104(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK11:       omp_offload.cont17:
+// CHECK11-NEXT:    [[TMP78:%.*]] = load i32, i32* [[M]], align 4
+// CHECK11-NEXT:    store i32 [[TMP78]], i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK11-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP79]], i32* [[N_CASTED19]], align 4
+// CHECK11-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_CASTED19]], align 4
+// CHECK11-NEXT:    [[TMP81:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK11-NEXT:    store i32 [[TMP81]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK11-NEXT:    [[TMP82:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK11-NEXT:    [[TMP83:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP84:%.*]] = sext i32 [[TMP83]] to i64
+// CHECK11-NEXT:    [[TMP85:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES23]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP85]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
+// CHECK11-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP66]], i32* [[TMP87]], align 4
-// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK11-NEXT:    store i32 [[TMP80]], i32* [[TMP87]], align 4
+// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP66]], i32* [[TMP89]], align 4
-// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK11-NEXT:    store i32 [[TMP80]], i32* [[TMP89]], align 4
+// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
 // CHECK11-NEXT:    store i8* null, i8** [[TMP90]], align 4
-// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP94:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP94]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP95]], 0
-// CHECK11-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
-// CHECK11-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
-// CHECK11-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK11-NEXT:    [[TMP96:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK11-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP96]], 1
-// CHECK11-NEXT:    [[TMP97:%.*]] = zext i32 [[ADD29]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP97]])
-// CHECK11-NEXT:    [[TMP98:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108.region_id, i32 4, i8** [[TMP91]], i8** [[TMP92]], i64* [[TMP93]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP99:%.*]] = icmp ne i32 [[TMP98]], 0
-// CHECK11-NEXT:    br i1 [[TMP99]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK11:       omp_offload.failed30:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108(i32 [[TMP64]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP66]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK11:       omp_offload.cont31:
-// CHECK11-NEXT:    [[TMP100:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP100]])
+// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP92]], align 4
+// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP94]], align 4
+// CHECK11-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP95]], align 4
+// CHECK11-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 4
+// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 4
+// CHECK11-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP84]], i64* [[TMP100]], align 4
+// CHECK11-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP101]], align 4
+// CHECK11-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP82]], i32* [[TMP103]], align 4
+// CHECK11-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP82]], i32* [[TMP105]], align 4
+// CHECK11-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP106]], align 4
+// CHECK11-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP110:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP110]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP111]], 0
+// CHECK11-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
+// CHECK11-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
+// CHECK11-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK11-NEXT:    [[TMP112:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK11-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP112]], 1
+// CHECK11-NEXT:    [[TMP113:%.*]] = zext i32 [[ADD30]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP113]])
+// CHECK11-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP114]], align 4
+// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP115]], align 4
+// CHECK11-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK11-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP108]], i8*** [[TMP117]], align 4
+// CHECK11-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP109]], i64** [[TMP118]], align 4
+// CHECK11-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP119]], align 4
+// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK11-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP121]], align 4
+// CHECK11-NEXT:    [[TMP122:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK11-NEXT:    [[TMP123:%.*]] = icmp ne i32 [[TMP122]], 0
+// CHECK11-NEXT:    br i1 [[TMP123]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK11:       omp_offload.failed32:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108(i32 [[TMP80]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP82]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK11:       omp_offload.cont33:
+// CHECK11-NEXT:    [[TMP124:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP124]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP101:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP101]])
-// CHECK11-NEXT:    [[TMP102:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP102]]
+// CHECK11-NEXT:    [[TMP125:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP125]])
+// CHECK11-NEXT:    [[TMP126:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP126]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100
@@ -4227,10 +4482,10 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -4244,62 +4499,113 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l81.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l81([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK11-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l85.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK11-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK11:       omp_offload.failed5:
+// CHECK11-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l85.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK11-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK11-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK11:       omp_offload.failed6:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l85([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK11:       omp_offload.cont6:
-// CHECK11-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK11-NEXT:    store i32 [[TMP18]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK11-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK11-NEXT:    store i32 [[TMP19]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK11-NEXT:    [[TMP20:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP22]], align 4
-// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP24]], align 4
-// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP25]], align 4
-// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP20]], i32* [[TMP27]], align 4
-// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP20]], i32* [[TMP29]], align 4
-// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP30]], align 4
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK11:       omp_offload.cont7:
+// CHECK11-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK11-NEXT:    store i32 [[TMP34]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK11-NEXT:    store i32 [[TMP35]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP36]], i32* [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP36]], i32* [[TMP45]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, i32 2, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK11-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK11:       omp_offload.failed11:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]], i32 [[TMP20]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK11:       omp_offload.cont12:
+// CHECK11-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 2, i32* [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP48]], i8*** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP55]], align 4
+// CHECK11-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP56]], align 4
+// CHECK11-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK11-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
+// CHECK11-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK11:       omp_offload.failed13:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]], i32 [[TMP36]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK11:       omp_offload.cont14:
 // CHECK11-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_firstprivate_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_firstprivate_codegen.cpp
index 6b38c2f..014ac7a 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_firstprivate_codegen.cpp
@@ -373,9 +373,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l122.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK1-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l122.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK1-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l122([2 x i32]* @vec, i64 [[TMP1]], [2 x %struct.S]* @s_arr, %struct.S* @var, i64 [[TMP3]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -798,27 +815,44 @@ int main() {
 // CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK1-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i64** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK1-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81([2 x i32]* [[VEC]], i64 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP30]]
+// CHECK1-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN2StC2Ev
@@ -1461,9 +1495,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l122.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK3-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l122.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK3-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l122([2 x i32]* @vec, i32 [[TMP1]], [2 x %struct.S]* @s_arr, %struct.S* @var, i32 [[TMP3]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1869,27 +1920,44 @@ int main() {
 // CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK3-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i64** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK3-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81([2 x i32]* [[VEC]], i32 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP30]]
+// CHECK3-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN2StC2Ev

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_if_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_if_codegen.cpp
index 4d35bcc..9b2b579 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_if_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_if_codegen.cpp
@@ -132,21 +132,55 @@ int main() {
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47(i64 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l53.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK1-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP19]], align 4
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP20]], align 4
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP22]], align 8
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l53.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
+// CHECK1-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l53() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret void
 //
 //
@@ -491,56 +525,90 @@ int main() {
 // CHECK1-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l78.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l78.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l78() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85() #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* @Arg, align 4
-// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP2]], 0
+// CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP10]], 0
 // CHECK1-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP3:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP3]] to i1
+// CHECK1-NEXT:    [[TMP11:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP11]] to i1
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i8*
 // CHECK1-NEXT:    [[FROMBOOL2:%.*]] = zext i1 [[TOBOOL1]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL2]], i8* [[CONV]], align 1
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP5]] to i1
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP13]] to i1
 // CHECK1-NEXT:    br i1 [[TOBOOL3]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK1:       omp_if.then:
-// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP7]], align 8
-// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP9]], align 8
-// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP13:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP13]] to i1
-// CHECK1-NEXT:    [[TMP14:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP21]] to i1
+// CHECK1-NEXT:    [[TMP22:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.10, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.11, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP14]])
-// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
-// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK1:       omp_offload.failed6:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP4]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK1:       omp_offload.cont7:
+// CHECK1-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP19]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.10, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.11, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP22]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK1-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
+// CHECK1-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP12]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK1:       omp_offload.cont8:
 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK1:       omp_if.else:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP4]]) #[[ATTR2]]
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP12]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* @Arg, align 4
-// CHECK1-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP17]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP33]])
 // CHECK1-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -1045,45 +1113,79 @@ int main() {
 // CHECK1-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l66() #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
-// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP2]], 0
+// CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP10]], 0
 // CHECK1-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP3:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP3]] to i1
+// CHECK1-NEXT:    [[TMP11:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP11]] to i1
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i8*
 // CHECK1-NEXT:    [[FROMBOOL2:%.*]] = zext i1 [[TOBOOL1]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL2]], i8* [[CONV]], align 1
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP6]], align 8
-// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP8]], align 8
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP9]], align 8
-// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP12]] to i1
-// CHECK1-NEXT:    [[TMP13:%.*]] = select i1 [[TOBOOL3]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP14:%.*]] = bitcast i8** [[TMP13]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP20]] to i1
+// CHECK1-NEXT:    [[TMP21:%.*]] = select i1 [[TOBOOL3]], i32 0, i32 1
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.18, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.19, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP13]])
-// CHECK1-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK1-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK1:       omp_offload.failed5:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70(i64 [[TMP4]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK1:       omp_offload.cont6:
+// CHECK1-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP18]], i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP19]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.18, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.19, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP21]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK1-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK1-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK1:       omp_offload.failed6:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70(i64 [[TMP12]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK1:       omp_offload.cont7:
 // CHECK1-NEXT:    ret i32 0
 //
 //
@@ -1606,21 +1708,55 @@ int main() {
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l45.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK3-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l45.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK3-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l45(i64 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l53.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK3-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK3:       omp_offload.failed2:
+// CHECK3-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP19]], align 4
+// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP21]], align 8
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP22]], align 8
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP23]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP24]], align 8
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP25]], align 8
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP26]], align 8
+// CHECK3-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l53.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK3-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
+// CHECK3-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK3:       omp_offload.failed3:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l53() #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK3:       omp_offload.cont3:
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK3:       omp_offload.cont4:
 // CHECK3-NEXT:    ret void
 //
 //
@@ -1965,56 +2101,90 @@ int main() {
 // CHECK3-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l78.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l78.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l78() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85() #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP2:%.*]] = load i32, i32* @Arg, align 4
-// CHECK3-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP2]], 0
+// CHECK3-NEXT:    [[TMP10:%.*]] = load i32, i32* @Arg, align 4
+// CHECK3-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP10]], 0
 // CHECK3-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK3-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK3-NEXT:    [[TMP3:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK3-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP3]] to i1
+// CHECK3-NEXT:    [[TMP11:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK3-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP11]] to i1
 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i8*
 // CHECK3-NEXT:    [[FROMBOOL2:%.*]] = zext i1 [[TOBOOL1]] to i8
 // CHECK3-NEXT:    store i8 [[FROMBOOL2]], i8* [[CONV]], align 1
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK3-NEXT:    [[TMP5:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK3-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP5]] to i1
+// CHECK3-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK3-NEXT:    [[TMP13:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK3-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP13]] to i1
 // CHECK3-NEXT:    br i1 [[TOBOOL3]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK3:       omp_if.then:
-// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK3-NEXT:    store i64 [[TMP4]], i64* [[TMP7]], align 8
-// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK3-NEXT:    store i64 [[TMP4]], i64* [[TMP9]], align 8
-// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP13:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK3-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP13]] to i1
-// CHECK3-NEXT:    [[TMP14:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i64*
+// CHECK3-NEXT:    store i64 [[TMP12]], i64* [[TMP15]], align 8
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
+// CHECK3-NEXT:    store i64 [[TMP12]], i64* [[TMP17]], align 8
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP18]], align 8
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP21:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK3-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP21]] to i1
+// CHECK3-NEXT:    [[TMP22:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP14]])
-// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
-// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK3:       omp_offload.failed6:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP4]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK3:       omp_offload.cont7:
+// CHECK3-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP19]], i8*** [[TMP25]], align 8
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP20]], i8*** [[TMP26]], align 8
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK3-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP22]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK3-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
+// CHECK3-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP12]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK3:       omp_offload.cont8:
 // CHECK3-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK3:       omp_if.else:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP4]]) #[[ATTR2]]
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP12]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP17:%.*]] = load i32, i32* @Arg, align 4
-// CHECK3-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP17]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = load i32, i32* @Arg, align 4
+// CHECK3-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP33]])
 // CHECK3-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -2753,45 +2923,79 @@ int main() {
 // CHECK3-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l66() #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
-// CHECK3-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP2]], 0
+// CHECK3-NEXT:    [[TMP10:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK3-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP10]], 0
 // CHECK3-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK3-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK3-NEXT:    [[TMP3:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK3-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP3]] to i1
+// CHECK3-NEXT:    [[TMP11:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK3-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP11]] to i1
 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i8*
 // CHECK3-NEXT:    [[FROMBOOL2:%.*]] = zext i1 [[TOBOOL1]] to i8
 // CHECK3-NEXT:    store i8 [[FROMBOOL2]], i8* [[CONV]], align 1
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
-// CHECK3-NEXT:    store i64 [[TMP4]], i64* [[TMP6]], align 8
-// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK3-NEXT:    store i64 [[TMP4]], i64* [[TMP8]], align 8
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP9]], align 8
-// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK3-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP12]] to i1
-// CHECK3-NEXT:    [[TMP13:%.*]] = select i1 [[TOBOOL3]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP14:%.*]] = bitcast i8** [[TMP13]] to i64*
+// CHECK3-NEXT:    store i64 [[TMP12]], i64* [[TMP14]], align 8
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK3-NEXT:    store i64 [[TMP12]], i64* [[TMP16]], align 8
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP17]], align 8
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP20:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK3-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP20]] to i1
+// CHECK3-NEXT:    [[TMP21:%.*]] = select i1 [[TOBOOL3]], i32 0, i32 1
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.19, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.20, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP13]])
-// CHECK3-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK3-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK3:       omp_offload.failed5:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70(i64 [[TMP4]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK3:       omp_offload.cont6:
+// CHECK3-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP18]], i8*** [[TMP24]], align 8
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP19]], i8*** [[TMP25]], align 8
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.19, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.20, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK3-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP21]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK3-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK3-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK3:       omp_offload.failed6:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70(i64 [[TMP12]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK3:       omp_offload.cont7:
 // CHECK3-NEXT:    ret i32 0
 //
 //
@@ -3907,21 +4111,55 @@ int main() {
 // CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK9-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK9-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47(i64 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l53.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK9-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK9:       omp_offload.failed2:
+// CHECK9-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP19]], align 4
+// CHECK9-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP20]], align 4
+// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP21]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP22]], align 8
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP23]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP24]], align 8
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP25]], align 8
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l53.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK9-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
+// CHECK9-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK9:       omp_offload.failed3:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l53() #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK9:       omp_offload.cont3:
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK9:       omp_offload.cont4:
 // CHECK9-NEXT:    ret void
 //
 //
@@ -4266,56 +4504,90 @@ int main() {
 // CHECK9-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l78.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l78.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l78() #[[ATTR2]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85() #[[ATTR2]]
-// CHECK9-NEXT:    [[TMP2:%.*]] = load i32, i32* @Arg, align 4
-// CHECK9-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP2]], 0
+// CHECK9-NEXT:    [[TMP10:%.*]] = load i32, i32* @Arg, align 4
+// CHECK9-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP10]], 0
 // CHECK9-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK9-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK9-NEXT:    [[TMP3:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK9-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP3]] to i1
+// CHECK9-NEXT:    [[TMP11:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK9-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP11]] to i1
 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i8*
 // CHECK9-NEXT:    [[FROMBOOL2:%.*]] = zext i1 [[TOBOOL1]] to i8
 // CHECK9-NEXT:    store i8 [[FROMBOOL2]], i8* [[CONV]], align 1
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK9-NEXT:    [[TMP5:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK9-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP5]] to i1
+// CHECK9-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK9-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP13]] to i1
 // CHECK9-NEXT:    br i1 [[TOBOOL3]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK9:       omp_if.then:
-// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP4]], i64* [[TMP7]], align 8
-// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP4]], i64* [[TMP9]], align 8
-// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP13:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK9-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP13]] to i1
-// CHECK9-NEXT:    [[TMP14:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP12]], i64* [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP12]], i64* [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP18]], align 8
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP21:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK9-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP21]] to i1
+// CHECK9-NEXT:    [[TMP22:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.10, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.11, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP14]])
-// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
-// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK9:       omp_offload.failed6:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP4]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK9:       omp_offload.cont7:
+// CHECK9-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP19]], i8*** [[TMP25]], align 8
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP20]], i8*** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.10, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.11, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP22]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK9-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
+// CHECK9-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK9:       omp_offload.failed7:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP12]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK9:       omp_offload.cont8:
 // CHECK9-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK9:       omp_if.else:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP4]]) #[[ATTR2]]
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP12]]) #[[ATTR2]]
 // CHECK9-NEXT:    br label [[OMP_IF_END]]
 // CHECK9:       omp_if.end:
-// CHECK9-NEXT:    [[TMP17:%.*]] = load i32, i32* @Arg, align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP17]])
+// CHECK9-NEXT:    [[TMP33:%.*]] = load i32, i32* @Arg, align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP33]])
 // CHECK9-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -4820,45 +5092,79 @@ int main() {
 // CHECK9-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62() #[[ATTR2]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l66() #[[ATTR2]]
-// CHECK9-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
-// CHECK9-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP2]], 0
+// CHECK9-NEXT:    [[TMP10:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK9-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP10]], 0
 // CHECK9-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK9-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK9-NEXT:    [[TMP3:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK9-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP3]] to i1
+// CHECK9-NEXT:    [[TMP11:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK9-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP11]] to i1
 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i8*
 // CHECK9-NEXT:    [[FROMBOOL2:%.*]] = zext i1 [[TOBOOL1]] to i8
 // CHECK9-NEXT:    store i8 [[FROMBOOL2]], i8* [[CONV]], align 1
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP4]], i64* [[TMP6]], align 8
-// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP4]], i64* [[TMP8]], align 8
-// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP9]], align 8
-// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK9-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP12]] to i1
-// CHECK9-NEXT:    [[TMP13:%.*]] = select i1 [[TOBOOL3]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP14:%.*]] = bitcast i8** [[TMP13]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP12]], i64* [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP12]], i64* [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP20:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK9-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP20]] to i1
+// CHECK9-NEXT:    [[TMP21:%.*]] = select i1 [[TOBOOL3]], i32 0, i32 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.18, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.19, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP13]])
-// CHECK9-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK9-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK9:       omp_offload.failed5:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70(i64 [[TMP4]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK9:       omp_offload.cont6:
+// CHECK9-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP18]], i8*** [[TMP24]], align 8
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP19]], i8*** [[TMP25]], align 8
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.18, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.19, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP21]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK9:       omp_offload.failed6:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70(i64 [[TMP12]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK9:       omp_offload.cont7:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -5381,21 +5687,55 @@ int main() {
 // CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l45.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK11-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK11-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK11-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l45.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK11-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l45(i64 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l53.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK11-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK11:       omp_offload.failed2:
+// CHECK11-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP19]], align 4
+// CHECK11-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP20]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP21]], align 8
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP22]], align 8
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP23]], align 8
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP24]], align 8
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP25]], align 8
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP26]], align 8
+// CHECK11-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l53.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK11-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
+// CHECK11-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK11:       omp_offload.failed3:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l53() #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK11:       omp_offload.cont3:
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK11:       omp_offload.cont4:
 // CHECK11-NEXT:    ret void
 //
 //
@@ -5740,56 +6080,90 @@ int main() {
 // CHECK11-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l78.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK11-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK11-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l78.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK11-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l78() #[[ATTR2]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l85() #[[ATTR2]]
-// CHECK11-NEXT:    [[TMP2:%.*]] = load i32, i32* @Arg, align 4
-// CHECK11-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP2]], 0
+// CHECK11-NEXT:    [[TMP10:%.*]] = load i32, i32* @Arg, align 4
+// CHECK11-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP10]], 0
 // CHECK11-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK11-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK11-NEXT:    [[TMP3:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK11-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP3]] to i1
+// CHECK11-NEXT:    [[TMP11:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK11-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP11]] to i1
 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i8*
 // CHECK11-NEXT:    [[FROMBOOL2:%.*]] = zext i1 [[TOBOOL1]] to i8
 // CHECK11-NEXT:    store i8 [[FROMBOOL2]], i8* [[CONV]], align 1
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK11-NEXT:    [[TMP5:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK11-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP5]] to i1
+// CHECK11-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK11-NEXT:    [[TMP13:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK11-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP13]] to i1
 // CHECK11-NEXT:    br i1 [[TOBOOL3]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK11:       omp_if.then:
-// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK11-NEXT:    store i64 [[TMP4]], i64* [[TMP7]], align 8
-// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK11-NEXT:    store i64 [[TMP4]], i64* [[TMP9]], align 8
-// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP13:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK11-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP13]] to i1
-// CHECK11-NEXT:    [[TMP14:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i64*
+// CHECK11-NEXT:    store i64 [[TMP12]], i64* [[TMP15]], align 8
+// CHECK11-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
+// CHECK11-NEXT:    store i64 [[TMP12]], i64* [[TMP17]], align 8
+// CHECK11-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP18]], align 8
+// CHECK11-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP21:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK11-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP21]] to i1
+// CHECK11-NEXT:    [[TMP22:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP14]])
-// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
-// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK11:       omp_offload.failed6:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP4]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK11:       omp_offload.cont7:
+// CHECK11-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP19]], i8*** [[TMP25]], align 8
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP20]], i8*** [[TMP26]], align 8
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK11-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP22]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK11-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
+// CHECK11-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK11:       omp_offload.failed7:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP12]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK11:       omp_offload.cont8:
 // CHECK11-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK11:       omp_if.else:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP4]]) #[[ATTR2]]
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP12]]) #[[ATTR2]]
 // CHECK11-NEXT:    br label [[OMP_IF_END]]
 // CHECK11:       omp_if.end:
-// CHECK11-NEXT:    [[TMP17:%.*]] = load i32, i32* @Arg, align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP17]])
+// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* @Arg, align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP33]])
 // CHECK11-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -6528,45 +6902,79 @@ int main() {
 // CHECK11-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK11-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK11-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK11-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62() #[[ATTR2]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l66() #[[ATTR2]]
-// CHECK11-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
-// CHECK11-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP2]], 0
+// CHECK11-NEXT:    [[TMP10:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK11-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP10]], 0
 // CHECK11-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK11-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK11-NEXT:    [[TMP3:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK11-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP3]] to i1
+// CHECK11-NEXT:    [[TMP11:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK11-NEXT:    [[TOBOOL1:%.*]] = trunc i8 [[TMP11]] to i1
 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i8*
 // CHECK11-NEXT:    [[FROMBOOL2:%.*]] = zext i1 [[TOBOOL1]] to i8
 // CHECK11-NEXT:    store i8 [[FROMBOOL2]], i8* [[CONV]], align 1
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
-// CHECK11-NEXT:    store i64 [[TMP4]], i64* [[TMP6]], align 8
-// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK11-NEXT:    store i64 [[TMP4]], i64* [[TMP8]], align 8
-// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP9]], align 8
-// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK11-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP12]] to i1
-// CHECK11-NEXT:    [[TMP13:%.*]] = select i1 [[TOBOOL3]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP12:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP14:%.*]] = bitcast i8** [[TMP13]] to i64*
+// CHECK11-NEXT:    store i64 [[TMP12]], i64* [[TMP14]], align 8
+// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK11-NEXT:    store i64 [[TMP12]], i64* [[TMP16]], align 8
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP17]], align 8
+// CHECK11-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP20:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK11-NEXT:    [[TOBOOL3:%.*]] = trunc i8 [[TMP20]] to i1
+// CHECK11-NEXT:    [[TMP21:%.*]] = select i1 [[TOBOOL3]], i32 0, i32 1
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.19, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.20, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP13]])
-// CHECK11-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK11-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK11:       omp_offload.failed5:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70(i64 [[TMP4]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK11:       omp_offload.cont6:
+// CHECK11-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP18]], i8*** [[TMP24]], align 8
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP19]], i8*** [[TMP25]], align 8
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.19, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.20, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK11-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP21]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK11-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK11-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK11:       omp_offload.failed6:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l70(i64 [[TMP12]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK11:       omp_offload.cont7:
 // CHECK11-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_lastprivate_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_lastprivate_codegen.cpp
index 8c0651f..ba454e9 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_lastprivate_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_lastprivate_codegen.cpp
@@ -862,9 +862,26 @@ int main() {
 // CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK5-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, i32 5, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK5-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK5-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 5, i32* [[TMP35]], align 4
+// CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP32]], i8*** [[TMP36]], align 8
+// CHECK5-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP33]], i8*** [[TMP37]], align 8
+// CHECK5-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 8
+// CHECK5-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 8
+// CHECK5-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP40]], align 8
+// CHECK5-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP41]], align 8
+// CHECK5-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK5-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106([2 x i32]* [[VEC]], i64 [[TMP3]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -872,18 +889,18 @@ int main() {
 // CHECK5-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK5-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK5-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK5-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK5-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK5:       arraydestroy.body:
-// CHECK5-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP36]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK5-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP44]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK5-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK5-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK5-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK5-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK5:       arraydestroy.done3:
 // CHECK5-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK5-NEXT:    [[TMP37:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP37]]
+// CHECK5-NEXT:    [[TMP45:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP45]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1355,27 +1372,44 @@ int main() {
 // CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK5-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK5-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 8
+// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP32]], align 8
+// CHECK5-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK5-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP34]], align 8
+// CHECK5-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK5-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50([2 x i32]* [[VEC]], i64 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
 // CHECK5-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK5-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK5-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK5-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK5:       arraydestroy.body:
-// CHECK5-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK5-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK5-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK5-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK5-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK5-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK5:       arraydestroy.done2:
 // CHECK5-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK5-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP30]]
+// CHECK5-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1913,9 +1947,26 @@ int main() {
 // CHECK7-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK7-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, i32 5, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK7-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK7-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 5, i32* [[TMP35]], align 4
+// CHECK7-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP32]], i8*** [[TMP36]], align 4
+// CHECK7-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP33]], i8*** [[TMP37]], align 4
+// CHECK7-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 4
+// CHECK7-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 4
+// CHECK7-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP40]], align 4
+// CHECK7-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP41]], align 4
+// CHECK7-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK7-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106([2 x i32]* [[VEC]], i32 [[TMP3]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1923,18 +1974,18 @@ int main() {
 // CHECK7-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK7-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK7-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK7-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK7-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK7:       arraydestroy.body:
-// CHECK7-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP36]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK7-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP44]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK7-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK7-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK7-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK7-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK7:       arraydestroy.done2:
 // CHECK7-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK7-NEXT:    [[TMP37:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK7-NEXT:    ret i32 [[TMP37]]
+// CHECK7-NEXT:    [[TMP45:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK7-NEXT:    ret i32 [[TMP45]]
 //
 //
 // CHECK7-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -2389,27 +2440,44 @@ int main() {
 // CHECK7-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK7-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK7-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK7-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK7-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 4
+// CHECK7-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK7-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP32]], align 4
+// CHECK7-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK7-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP34]], align 4
+// CHECK7-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK7-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l50([2 x i32]* [[VEC]], i32 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK7:       omp_offload.cont:
 // CHECK7-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK7-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK7-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK7-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK7:       arraydestroy.body:
-// CHECK7-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK7-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK7-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK7-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK7-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK7-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK7:       arraydestroy.done2:
 // CHECK7-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK7-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK7-NEXT:    ret i32 [[TMP30]]
+// CHECK7-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK7-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK7-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_private_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_private_codegen.cpp
index 1759798..08c9c21 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_private_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_private_codegen.cpp
@@ -300,9 +300,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l124.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l124.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l124() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -561,27 +578,44 @@ int main() {
 // CHECK1-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 8
 // CHECK1-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 8
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l80.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK1-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l80() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP4]]
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev
@@ -1021,9 +1055,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l124.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l124.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l124() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1276,27 +1327,44 @@ int main() {
 // CHECK3-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 4
 // CHECK3-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l80.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK3-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l80() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP4]]
+// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_proc_bind_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_proc_bind_codegen.cpp
index 3886765..c420768 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_proc_bind_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_proc_bind_codegen.cpp
@@ -56,21 +56,55 @@ int main() {
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l36.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l36() #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l38.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l38.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l38() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK1-NEXT:    ret i32 [[CALL]]
 //
@@ -384,9 +418,26 @@ int main() {
 // CHECK1-NEXT:  entry:
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l30.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l30.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l30() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_reduction_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_reduction_codegen.cpp
index 2ff2c5c..b1e93fe 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_reduction_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_reduction_codegen.cpp
@@ -110,9 +110,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l66.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l66.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l66(i32* @_ZZ4mainE5sivar) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -404,9 +421,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
-// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
+// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i32* [[T_VAR]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -701,9 +735,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l66.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l66.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l66(i32* @_ZZ4mainE5sivar) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -991,9 +1042,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
-// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
+// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i32* [[T_VAR]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_schedule_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_schedule_codegen.cpp
index e6a2889..9ede1cf 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_schedule_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_schedule_codegen.cpp
@@ -214,18 +214,18 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -240,93 +240,178 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK1:       omp_offload.failed16:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK1:       omp_offload.cont15:
-// CHECK1-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 8
-// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 8
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP31]], align 8
-// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK1:       omp_offload.cont17:
+// CHECK1-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 8
+// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 8
+// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP55]], align 8
+// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK1-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK1:       omp_offload.failed21:
+// CHECK1-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK1-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK1-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 8
+// CHECK1-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 8
+// CHECK1-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 8
+// CHECK1-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 8
+// CHECK1-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP64]], align 8
+// CHECK1-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP65]], align 8
+// CHECK1-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK1-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK1-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK1:       omp_offload.failed24:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK1:       omp_offload.cont22:
-// CHECK1-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 8
-// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 8
-// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK1:       omp_offload.cont25:
+// CHECK1-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 8
+// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP72]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK1-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK1:       omp_offload.failed28:
+// CHECK1-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK1-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 8
+// CHECK1-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 8
+// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 8
+// CHECK1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 8
+// CHECK1-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP81]], align 8
+// CHECK1-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP82]], align 8
+// CHECK1-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK1-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK1-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK1:       omp_offload.failed32:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK1:       omp_offload.cont29:
-// CHECK1-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP45]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK1:       omp_offload.cont33:
+// CHECK1-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i64 0, i64 0
+// CHECK1-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36
@@ -1198,18 +1283,18 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -1224,93 +1309,178 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK3:       omp_offload.failed8:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK3:       omp_offload.cont8:
-// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK3:       omp_offload.cont9:
+// CHECK3-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK3-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK3:       omp_offload.failed14:
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK3:       omp_offload.cont15:
-// CHECK3-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 4
-// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 4
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP31]], align 4
-// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 4
+// CHECK3-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 4
+// CHECK3-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP55]], align 4
+// CHECK3-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK3-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK3:       omp_offload.failed21:
+// CHECK3-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK3-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK3-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 4
+// CHECK3-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 4
+// CHECK3-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 4
+// CHECK3-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 4
+// CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP64]], align 4
+// CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP65]], align 4
+// CHECK3-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK3-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK3-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK3:       omp_offload.failed24:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK3:       omp_offload.cont22:
-// CHECK3-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 4
-// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 4
-// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP40]], align 4
-// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK3:       omp_offload.cont25:
+// CHECK3-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 4
+// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP72]], align 4
+// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK3-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK3:       omp_offload.failed28:
+// CHECK3-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK3-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK3-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 4
+// CHECK3-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 4
+// CHECK3-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 4
+// CHECK3-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 4
+// CHECK3-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP81]], align 4
+// CHECK3-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP82]], align 4
+// CHECK3-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK3-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK3-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK3:       omp_offload.failed32:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK3:       omp_offload.cont29:
-// CHECK3-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP45]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK3:       omp_offload.cont33:
+// CHECK3-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36
@@ -2155,18 +2325,18 @@ int main (int argc, char **argv) {
 // CHECK5-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK5-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK5-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK5-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -2181,93 +2351,178 @@ int main (int argc, char **argv) {
 // CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK5-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK5-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK5-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK5-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK5-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
 // CHECK5-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK5-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK5-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK5-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK5-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK5:       omp_offload.failed7:
+// CHECK5-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK5-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK5-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK5-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK5:       omp_offload.failed8:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK5:       omp_offload.cont8:
-// CHECK5-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK5-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK5-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK5-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK5:       omp_offload.cont9:
+// CHECK5-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK5-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK5-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK5-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK5:       omp_offload.failed14:
+// CHECK5-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK5-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK5-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK5-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK5-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK5-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK5-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK5-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK5-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK5-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK5-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK5:       omp_offload.failed16:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK5:       omp_offload.cont15:
-// CHECK5-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 8
-// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 8
-// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP31]], align 8
-// CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK5:       omp_offload.cont17:
+// CHECK5-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 8
+// CHECK5-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 8
+// CHECK5-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP55]], align 8
+// CHECK5-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK5-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK5:       omp_offload.failed21:
+// CHECK5-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK5-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK5-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 8
+// CHECK5-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 8
+// CHECK5-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 8
+// CHECK5-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 8
+// CHECK5-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP64]], align 8
+// CHECK5-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP65]], align 8
+// CHECK5-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK5-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK5-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK5:       omp_offload.failed24:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK5:       omp_offload.cont22:
-// CHECK5-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 8
-// CHECK5-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 8
-// CHECK5-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK5-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK5:       omp_offload.cont25:
+// CHECK5-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 8
+// CHECK5-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 8
+// CHECK5-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP72]], align 8
+// CHECK5-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK5-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK5:       omp_offload.failed28:
+// CHECK5-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK5-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK5-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 8
+// CHECK5-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 8
+// CHECK5-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 8
+// CHECK5-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 8
+// CHECK5-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP81]], align 8
+// CHECK5-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP82]], align 8
+// CHECK5-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK5-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK5-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK5:       omp_offload.failed32:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK5:       omp_offload.cont29:
-// CHECK5-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i64 0, i64 0
-// CHECK5-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP45]]
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK5:       omp_offload.cont33:
+// CHECK5-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i64 0, i64 0
+// CHECK5-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36
@@ -3139,18 +3394,18 @@ int main (int argc, char **argv) {
 // CHECK7-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK7-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK7-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK7-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK7-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -3165,93 +3420,178 @@ int main (int argc, char **argv) {
 // CHECK7-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK7-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK7-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK7-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK7-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK7-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK7-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK7-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK7-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK7-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK7-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK7:       omp_offload.cont:
 // CHECK7-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK7-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK7-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK7-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK7-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK7-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK7-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK7-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK7-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK7-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK7:       omp_offload.failed7:
+// CHECK7-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK7-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK7-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK7-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK7-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK7-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK7-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK7-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK7-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK7:       omp_offload.failed8:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK7:       omp_offload.cont8:
-// CHECK7-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK7-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK7-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK7-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK7-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK7:       omp_offload.cont9:
+// CHECK7-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK7-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK7-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK7-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK7-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK7-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK7:       omp_offload.failed14:
+// CHECK7-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK7-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK7-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK7-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK7-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK7-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK7-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK7-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK7-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK7-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK7-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK7:       omp_offload.failed16:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l44(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK7:       omp_offload.cont15:
-// CHECK7-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 4
-// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK7-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 4
-// CHECK7-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP31]], align 4
-// CHECK7-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK7:       omp_offload.cont17:
+// CHECK7-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 4
+// CHECK7-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK7-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 4
+// CHECK7-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP55]], align 4
+// CHECK7-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK7-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK7:       omp_offload.failed21:
+// CHECK7-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK7-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK7-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 4
+// CHECK7-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 4
+// CHECK7-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 4
+// CHECK7-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 4
+// CHECK7-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP64]], align 4
+// CHECK7-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP65]], align 4
+// CHECK7-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK7-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK7-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK7:       omp_offload.failed24:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l49(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK7:       omp_offload.cont22:
-// CHECK7-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 4
-// CHECK7-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK7-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 4
-// CHECK7-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP40]], align 4
-// CHECK7-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK7:       omp_offload.cont25:
+// CHECK7-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 4
+// CHECK7-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK7-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 4
+// CHECK7-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP72]], align 4
+// CHECK7-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK7-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK7:       omp_offload.failed28:
+// CHECK7-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK7-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK7-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 4
+// CHECK7-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 4
+// CHECK7-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 4
+// CHECK7-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 4
+// CHECK7-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP81]], align 4
+// CHECK7-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP82]], align 4
+// CHECK7-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK7-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK7-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK7:       omp_offload.failed32:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l54(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK7:       omp_offload.cont29:
-// CHECK7-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK7-NEXT:    ret i32 [[TMP45]]
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK7:       omp_offload.cont33:
+// CHECK7-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK7-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK7-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36
@@ -4487,34 +4827,34 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[_TMP9:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
-// CHECK13-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES26:%.*]] = alloca [4 x i64], align 8
+// CHECK13-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[N_CASTED35:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS37:%.*]] = alloca [3 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS38:%.*]] = alloca [3 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS39:%.*]] = alloca [3 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES40:%.*]] = alloca [3 x i64], align 8
-// CHECK13-NEXT:    [[_TMP41:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_42:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_43:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_50:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[N_CASTED51:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR__CASTED53:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS55:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS56:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS57:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES58:%.*]] = alloca [4 x i64], align 8
-// CHECK13-NEXT:    [[_TMP59:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_60:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_61:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_29:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[N_CASTED37:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS39:%.*]] = alloca [3 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS40:%.*]] = alloca [3 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS41:%.*]] = alloca [3 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES42:%.*]] = alloca [3 x i64], align 8
+// CHECK13-NEXT:    [[_TMP43:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_44:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_45:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_53:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[N_CASTED54:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR__CASTED56:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS58:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS59:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS60:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES61:%.*]] = alloca [4 x i64], align 8
+// CHECK13-NEXT:    [[_TMP62:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_63:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_64:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK13-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK13-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -4573,264 +4913,349 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK13-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK13-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK13-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK13-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK13-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK13-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK13-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK13-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK13-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK13-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK13-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK13-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       omp_offload.cont:
-// CHECK13-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK13-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK13-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK13-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
-// CHECK13-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK13-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK13-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK13-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK13-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK13-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK13-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK13-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK13-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK13-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK13-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK13-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK13-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK13-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK13-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK13-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
+// CHECK13-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK13-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK13-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK13-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK13-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK13-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK13-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK13-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK13-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK13-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK13-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK13-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK13-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK13-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK13-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK13-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK13-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK13-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP58]])
-// CHECK13-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK13-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK13:       omp_offload.failed16:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK13:       omp_offload.cont17:
-// CHECK13-NEXT:    [[TMP61:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    store i32 [[TMP61]], i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK13-NEXT:    [[TMP62:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP62]], i32* [[CONV20]], align 4
-// CHECK13-NEXT:    [[TMP63:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK13-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK13-NEXT:    [[CONV21:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP64]], i32* [[CONV21]], align 4
-// CHECK13-NEXT:    [[TMP65:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK13-NEXT:    [[TMP66:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP67:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP67]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
-// CHECK13-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP63]], i64* [[TMP69]], align 8
-// CHECK13-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP63]], i64* [[TMP71]], align 8
-// CHECK13-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP72]], align 8
-// CHECK13-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP74]], align 8
-// CHECK13-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP76]], align 8
-// CHECK13-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP77]], align 8
-// CHECK13-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP79]], align 8
-// CHECK13-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 8
-// CHECK13-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 2
-// CHECK13-NEXT:    store i64 [[TMP66]], i64* [[TMP82]], align 8
-// CHECK13-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP83]], align 8
-// CHECK13-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK13-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK13-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP66]])
+// CHECK13-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK13-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK13-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK13-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK13-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK13-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK13-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK13-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK13-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK13-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK13-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK13:       omp_offload.failed17:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK13:       omp_offload.cont18:
+// CHECK13-NEXT:    [[TMP77:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    store i32 [[TMP77]], i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK13-NEXT:    [[TMP78:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP78]], i32* [[CONV21]], align 4
+// CHECK13-NEXT:    [[TMP79:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK13-NEXT:    [[TMP80:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK13-NEXT:    [[CONV22:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP80]], i32* [[CONV22]], align 4
+// CHECK13-NEXT:    [[TMP81:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK13-NEXT:    [[TMP82:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP83:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES26]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP83]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
+// CHECK13-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
 // CHECK13-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP65]], i64* [[TMP85]], align 8
-// CHECK13-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK13-NEXT:    store i64 [[TMP79]], i64* [[TMP85]], align 8
+// CHECK13-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
 // CHECK13-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP65]], i64* [[TMP87]], align 8
-// CHECK13-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK13-NEXT:    store i64 [[TMP79]], i64* [[TMP87]], align 8
+// CHECK13-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 0
 // CHECK13-NEXT:    store i8* null, i8** [[TMP88]], align 8
-// CHECK13-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP92:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP92]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK13-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK13-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP93]], 0
-// CHECK13-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
-// CHECK13-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
-// CHECK13-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK13-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK13-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP94]], 1
-// CHECK13-NEXT:    [[TMP95:%.*]] = zext i32 [[ADD32]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP95]])
-// CHECK13-NEXT:    [[TMP96:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, i32 4, i8** [[TMP89]], i8** [[TMP90]], i64* [[TMP91]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP97:%.*]] = icmp ne i32 [[TMP96]], 0
-// CHECK13-NEXT:    br i1 [[TMP97]], label [[OMP_OFFLOAD_FAILED33:%.*]], label [[OMP_OFFLOAD_CONT34:%.*]]
-// CHECK13:       omp_offload.failed33:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i64 [[TMP63]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP65]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT34]]
-// CHECK13:       omp_offload.cont34:
-// CHECK13-NEXT:    [[TMP98:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    [[CONV36:%.*]] = bitcast i64* [[N_CASTED35]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP98]], i32* [[CONV36]], align 4
-// CHECK13-NEXT:    [[TMP99:%.*]] = load i64, i64* [[N_CASTED35]], align 8
-// CHECK13-NEXT:    [[TMP100:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP101:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES40]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP101]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
-// CHECK13-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP90]], align 8
+// CHECK13-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP92]], align 8
+// CHECK13-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP93]], align 8
+// CHECK13-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP95]], align 8
+// CHECK13-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 8
+// CHECK13-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 2
+// CHECK13-NEXT:    store i64 [[TMP82]], i64* [[TMP98]], align 8
+// CHECK13-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP99]], align 8
+// CHECK13-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP81]], i64* [[TMP101]], align 8
+// CHECK13-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 3
 // CHECK13-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP99]], i64* [[TMP103]], align 8
-// CHECK13-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP99]], i64* [[TMP105]], align 8
-// CHECK13-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS39]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP106]], align 8
-// CHECK13-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP108]], align 8
-// CHECK13-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP110]], align 8
-// CHECK13-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS39]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP111]], align 8
-// CHECK13-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP113]], align 8
-// CHECK13-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP115]], align 8
-// CHECK13-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES40]], i32 0, i32 2
-// CHECK13-NEXT:    store i64 [[TMP100]], i64* [[TMP116]], align 8
-// CHECK13-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS39]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP117]], align 8
-// CHECK13-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES40]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP121:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP121]], i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK13-NEXT:    [[TMP122:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK13-NEXT:    [[SUB44:%.*]] = sub nsw i32 [[TMP122]], 0
-// CHECK13-NEXT:    [[DIV45:%.*]] = sdiv i32 [[SUB44]], 1
-// CHECK13-NEXT:    [[SUB46:%.*]] = sub nsw i32 [[DIV45]], 1
-// CHECK13-NEXT:    store i32 [[SUB46]], i32* [[DOTCAPTURE_EXPR_43]], align 4
-// CHECK13-NEXT:    [[TMP123:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_43]], align 4
-// CHECK13-NEXT:    [[ADD47:%.*]] = add nsw i32 [[TMP123]], 1
-// CHECK13-NEXT:    [[TMP124:%.*]] = zext i32 [[ADD47]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP124]])
-// CHECK13-NEXT:    [[TMP125:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, i32 3, i8** [[TMP118]], i8** [[TMP119]], i64* [[TMP120]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP126:%.*]] = icmp ne i32 [[TMP125]], 0
-// CHECK13-NEXT:    br i1 [[TMP126]], label [[OMP_OFFLOAD_FAILED48:%.*]], label [[OMP_OFFLOAD_CONT49:%.*]]
-// CHECK13:       omp_offload.failed48:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i64 [[TMP99]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT49]]
-// CHECK13:       omp_offload.cont49:
-// CHECK13-NEXT:    [[TMP127:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    store i32 [[TMP127]], i32* [[DOTCAPTURE_EXPR_50]], align 4
-// CHECK13-NEXT:    [[TMP128:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    [[CONV52:%.*]] = bitcast i64* [[N_CASTED51]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP128]], i32* [[CONV52]], align 4
-// CHECK13-NEXT:    [[TMP129:%.*]] = load i64, i64* [[N_CASTED51]], align 8
-// CHECK13-NEXT:    [[TMP130:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_50]], align 4
-// CHECK13-NEXT:    [[CONV54:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED53]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP130]], i32* [[CONV54]], align 4
-// CHECK13-NEXT:    [[TMP131:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED53]], align 8
-// CHECK13-NEXT:    [[TMP132:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP133:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES58]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP133]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
-// CHECK13-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP129]], i64* [[TMP135]], align 8
-// CHECK13-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP129]], i64* [[TMP137]], align 8
-// CHECK13-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP138]], align 8
-// CHECK13-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP140]], align 8
-// CHECK13-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP142]], align 8
-// CHECK13-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP143]], align 8
-// CHECK13-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP145:%.*]] = bitcast i8** [[TMP144]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP145]], align 8
-// CHECK13-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP147]], align 8
-// CHECK13-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES58]], i32 0, i32 2
-// CHECK13-NEXT:    store i64 [[TMP132]], i64* [[TMP148]], align 8
-// CHECK13-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP149]], align 8
-// CHECK13-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 3
-// CHECK13-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP131]], i64* [[TMP151]], align 8
-// CHECK13-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 3
-// CHECK13-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP131]], i64* [[TMP153]], align 8
-// CHECK13-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 3
-// CHECK13-NEXT:    store i8* null, i8** [[TMP154]], align 8
-// CHECK13-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES58]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP158:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP158]], i32* [[DOTCAPTURE_EXPR_60]], align 4
-// CHECK13-NEXT:    [[TMP159:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_60]], align 4
-// CHECK13-NEXT:    [[SUB62:%.*]] = sub nsw i32 [[TMP159]], 0
-// CHECK13-NEXT:    [[DIV63:%.*]] = sdiv i32 [[SUB62]], 1
-// CHECK13-NEXT:    [[SUB64:%.*]] = sub nsw i32 [[DIV63]], 1
-// CHECK13-NEXT:    store i32 [[SUB64]], i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK13-NEXT:    [[TMP160:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK13-NEXT:    [[ADD65:%.*]] = add nsw i32 [[TMP160]], 1
-// CHECK13-NEXT:    [[TMP161:%.*]] = zext i32 [[ADD65]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP161]])
-// CHECK13-NEXT:    [[TMP162:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, i32 4, i8** [[TMP155]], i8** [[TMP156]], i64* [[TMP157]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP163:%.*]] = icmp ne i32 [[TMP162]], 0
-// CHECK13-NEXT:    br i1 [[TMP163]], label [[OMP_OFFLOAD_FAILED66:%.*]], label [[OMP_OFFLOAD_CONT67:%.*]]
-// CHECK13:       omp_offload.failed66:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i64 [[TMP129]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP131]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT67]]
-// CHECK13:       omp_offload.cont67:
-// CHECK13-NEXT:    [[TMP164:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK13-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP164]])
+// CHECK13-NEXT:    store i64 [[TMP81]], i64* [[TMP103]], align 8
+// CHECK13-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 3
+// CHECK13-NEXT:    store i8* null, i8** [[TMP104]], align 8
+// CHECK13-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP108:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP108]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK13-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK13-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[TMP109]], 0
+// CHECK13-NEXT:    [[DIV31:%.*]] = sdiv i32 [[SUB30]], 1
+// CHECK13-NEXT:    [[SUB32:%.*]] = sub nsw i32 [[DIV31]], 1
+// CHECK13-NEXT:    store i32 [[SUB32]], i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK13-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK13-NEXT:    [[ADD33:%.*]] = add nsw i32 [[TMP110]], 1
+// CHECK13-NEXT:    [[TMP111:%.*]] = zext i32 [[ADD33]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP111]])
+// CHECK13-NEXT:    [[KERNEL_ARGS34:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP112]], align 4
+// CHECK13-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 4, i32* [[TMP113]], align 4
+// CHECK13-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK13-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 8
+// CHECK13-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP107]], i64** [[TMP116]], align 8
+// CHECK13-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP117]], align 8
+// CHECK13-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK13-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP119]], align 8
+// CHECK13-NEXT:    [[TMP120:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]])
+// CHECK13-NEXT:    [[TMP121:%.*]] = icmp ne i32 [[TMP120]], 0
+// CHECK13-NEXT:    br i1 [[TMP121]], label [[OMP_OFFLOAD_FAILED35:%.*]], label [[OMP_OFFLOAD_CONT36:%.*]]
+// CHECK13:       omp_offload.failed35:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i64 [[TMP79]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP81]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT36]]
+// CHECK13:       omp_offload.cont36:
+// CHECK13-NEXT:    [[TMP122:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[CONV38:%.*]] = bitcast i64* [[N_CASTED37]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP122]], i32* [[CONV38]], align 4
+// CHECK13-NEXT:    [[TMP123:%.*]] = load i64, i64* [[N_CASTED37]], align 8
+// CHECK13-NEXT:    [[TMP124:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP125:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES42]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP125]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
+// CHECK13-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP123]], i64* [[TMP127]], align 8
+// CHECK13-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP123]], i64* [[TMP129]], align 8
+// CHECK13-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS41]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP130]], align 8
+// CHECK13-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP132]], align 8
+// CHECK13-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP134]], align 8
+// CHECK13-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS41]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP135]], align 8
+// CHECK13-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP137]], align 8
+// CHECK13-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP139]], align 8
+// CHECK13-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES42]], i32 0, i32 2
+// CHECK13-NEXT:    store i64 [[TMP124]], i64* [[TMP140]], align 8
+// CHECK13-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS41]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP141]], align 8
+// CHECK13-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES42]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP145:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP145]], i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK13-NEXT:    [[TMP146:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK13-NEXT:    [[SUB46:%.*]] = sub nsw i32 [[TMP146]], 0
+// CHECK13-NEXT:    [[DIV47:%.*]] = sdiv i32 [[SUB46]], 1
+// CHECK13-NEXT:    [[SUB48:%.*]] = sub nsw i32 [[DIV47]], 1
+// CHECK13-NEXT:    store i32 [[SUB48]], i32* [[DOTCAPTURE_EXPR_45]], align 4
+// CHECK13-NEXT:    [[TMP147:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_45]], align 4
+// CHECK13-NEXT:    [[ADD49:%.*]] = add nsw i32 [[TMP147]], 1
+// CHECK13-NEXT:    [[TMP148:%.*]] = zext i32 [[ADD49]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP148]])
+// CHECK13-NEXT:    [[KERNEL_ARGS50:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP149]], align 4
+// CHECK13-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 3, i32* [[TMP150]], align 4
+// CHECK13-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP142]], i8*** [[TMP151]], align 8
+// CHECK13-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP143]], i8*** [[TMP152]], align 8
+// CHECK13-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP144]], i64** [[TMP153]], align 8
+// CHECK13-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP154]], align 8
+// CHECK13-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP155]], align 8
+// CHECK13-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP156]], align 8
+// CHECK13-NEXT:    [[TMP157:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]])
+// CHECK13-NEXT:    [[TMP158:%.*]] = icmp ne i32 [[TMP157]], 0
+// CHECK13-NEXT:    br i1 [[TMP158]], label [[OMP_OFFLOAD_FAILED51:%.*]], label [[OMP_OFFLOAD_CONT52:%.*]]
+// CHECK13:       omp_offload.failed51:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i64 [[TMP123]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT52]]
+// CHECK13:       omp_offload.cont52:
+// CHECK13-NEXT:    [[TMP159:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    store i32 [[TMP159]], i32* [[DOTCAPTURE_EXPR_53]], align 4
+// CHECK13-NEXT:    [[TMP160:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[CONV55:%.*]] = bitcast i64* [[N_CASTED54]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP160]], i32* [[CONV55]], align 4
+// CHECK13-NEXT:    [[TMP161:%.*]] = load i64, i64* [[N_CASTED54]], align 8
+// CHECK13-NEXT:    [[TMP162:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_53]], align 4
+// CHECK13-NEXT:    [[CONV57:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED56]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP162]], i32* [[CONV57]], align 4
+// CHECK13-NEXT:    [[TMP163:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED56]], align 8
+// CHECK13-NEXT:    [[TMP164:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP165:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES61]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP165]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
+// CHECK13-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP161]], i64* [[TMP167]], align 8
+// CHECK13-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP161]], i64* [[TMP169]], align 8
+// CHECK13-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP170]], align 8
+// CHECK13-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP172:%.*]] = bitcast i8** [[TMP171]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP172]], align 8
+// CHECK13-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP174:%.*]] = bitcast i8** [[TMP173]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP174]], align 8
+// CHECK13-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP175]], align 8
+// CHECK13-NEXT:    [[TMP176:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP177:%.*]] = bitcast i8** [[TMP176]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP177]], align 8
+// CHECK13-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP179]], align 8
+// CHECK13-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES61]], i32 0, i32 2
+// CHECK13-NEXT:    store i64 [[TMP164]], i64* [[TMP180]], align 8
+// CHECK13-NEXT:    [[TMP181:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP181]], align 8
+// CHECK13-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP183:%.*]] = bitcast i8** [[TMP182]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP163]], i64* [[TMP183]], align 8
+// CHECK13-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP185:%.*]] = bitcast i8** [[TMP184]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP163]], i64* [[TMP185]], align 8
+// CHECK13-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 3
+// CHECK13-NEXT:    store i8* null, i8** [[TMP186]], align 8
+// CHECK13-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES61]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP190:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP190]], i32* [[DOTCAPTURE_EXPR_63]], align 4
+// CHECK13-NEXT:    [[TMP191:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_63]], align 4
+// CHECK13-NEXT:    [[SUB65:%.*]] = sub nsw i32 [[TMP191]], 0
+// CHECK13-NEXT:    [[DIV66:%.*]] = sdiv i32 [[SUB65]], 1
+// CHECK13-NEXT:    [[SUB67:%.*]] = sub nsw i32 [[DIV66]], 1
+// CHECK13-NEXT:    store i32 [[SUB67]], i32* [[DOTCAPTURE_EXPR_64]], align 4
+// CHECK13-NEXT:    [[TMP192:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_64]], align 4
+// CHECK13-NEXT:    [[ADD68:%.*]] = add nsw i32 [[TMP192]], 1
+// CHECK13-NEXT:    [[TMP193:%.*]] = zext i32 [[ADD68]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP193]])
+// CHECK13-NEXT:    [[KERNEL_ARGS69:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP194]], align 4
+// CHECK13-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 4, i32* [[TMP195]], align 4
+// CHECK13-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP187]], i8*** [[TMP196]], align 8
+// CHECK13-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP188]], i8*** [[TMP197]], align 8
+// CHECK13-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP189]], i64** [[TMP198]], align 8
+// CHECK13-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP199]], align 8
+// CHECK13-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP200]], align 8
+// CHECK13-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP201]], align 8
+// CHECK13-NEXT:    [[TMP202:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]])
+// CHECK13-NEXT:    [[TMP203:%.*]] = icmp ne i32 [[TMP202]], 0
+// CHECK13-NEXT:    br i1 [[TMP203]], label [[OMP_OFFLOAD_FAILED70:%.*]], label [[OMP_OFFLOAD_CONT71:%.*]]
+// CHECK13:       omp_offload.failed70:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i64 [[TMP161]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP163]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT71]]
+// CHECK13:       omp_offload.cont71:
+// CHECK13-NEXT:    [[TMP204:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK13-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP204]])
 // CHECK13-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK13-NEXT:    [[TMP165:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK13-NEXT:    call void @llvm.stackrestore(i8* [[TMP165]])
-// CHECK13-NEXT:    [[TMP166:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK13-NEXT:    ret i32 [[TMP166]]
+// CHECK13-NEXT:    [[TMP205:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK13-NEXT:    call void @llvm.stackrestore(i8* [[TMP205]])
+// CHECK13-NEXT:    [[TMP206:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK13-NEXT:    ret i32 [[TMP206]]
 //
 //
 // CHECK13-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139
@@ -6135,20 +6560,20 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 8
-// CHECK13-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR__CASTED20:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 8
+// CHECK13-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR__CASTED23:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS25:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS26:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS27:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[_TMP28:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK13-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK13-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -6162,113 +6587,198 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK13-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK13-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK13-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK13-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK13-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK13-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK13-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK13-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK13-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       omp_offload.cont:
-// CHECK13-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK13-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK13-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK13-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK13-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK13-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK13:       omp_offload.failed5:
+// CHECK13-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK13-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK13-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK13-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK13-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK13-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK13-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK13-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK13-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK13-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK13-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK13:       omp_offload.failed6:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK13:       omp_offload.cont6:
-// CHECK13-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    store i32 [[TMP18]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK13-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK13:       omp_offload.cont7:
+// CHECK13-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    store i32 [[TMP34]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK13-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
 // CHECK13-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP19]], i32* [[CONV]], align 4
-// CHECK13-NEXT:    [[TMP20:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK13-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP22]], align 8
-// CHECK13-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP24]], align 8
-// CHECK13-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP25]], align 8
-// CHECK13-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP20]], i64* [[TMP27]], align 8
-// CHECK13-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP20]], i64* [[TMP29]], align 8
-// CHECK13-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP30]], align 8
-// CHECK13-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, i32 2, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK13-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK13:       omp_offload.failed11:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i64 [[TMP20]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK13:       omp_offload.cont12:
-// CHECK13-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP36]], align 8
-// CHECK13-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 [[TMP35]], i32* [[CONV]], align 4
+// CHECK13-NEXT:    [[TMP36:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK13-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
 // CHECK13-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [10 x i32]**
 // CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP38]], align 8
-// CHECK13-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP39]], align 8
-// CHECK13-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP40]], align 8
+// CHECK13-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP41]], align 8
+// CHECK13-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP36]], i64* [[TMP43]], align 8
+// CHECK13-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP36]], i64* [[TMP45]], align 8
+// CHECK13-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP46]], align 8
+// CHECK13-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, i32 1, i8** [[TMP40]], i8** [[TMP41]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK13-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK13:       omp_offload.failed17:
+// CHECK13-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP49]], align 4
+// CHECK13-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 2, i32* [[TMP50]], align 4
+// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 8
+// CHECK13-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP48]], i8*** [[TMP52]], align 8
+// CHECK13-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP54]], align 8
+// CHECK13-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK13-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP56]], align 8
+// CHECK13-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK13-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
+// CHECK13-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK13:       omp_offload.failed13:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i64 [[TMP36]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK13:       omp_offload.cont14:
+// CHECK13-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP60]], align 8
+// CHECK13-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP62]], align 8
+// CHECK13-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP63]], align 8
+// CHECK13-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
+// CHECK13-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK13-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK13-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 8
+// CHECK13-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 8
+// CHECK13-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP70]], align 8
+// CHECK13-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP71]], align 8
+// CHECK13-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP72]], align 8
+// CHECK13-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK13-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK13-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK13-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK13:       omp_offload.failed20:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK13:       omp_offload.cont18:
-// CHECK13-NEXT:    [[TMP44:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    store i32 [[TMP44]], i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK13-NEXT:    [[TMP45:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK13-NEXT:    [[CONV21:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED20]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP45]], i32* [[CONV21]], align 4
-// CHECK13-NEXT:    [[TMP46:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED20]], align 8
-// CHECK13-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP48]], align 8
-// CHECK13-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP50]], align 8
-// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK13-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP46]], i64* [[TMP53]], align 8
-// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP46]], i64* [[TMP55]], align 8
-// CHECK13-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP56]], align 8
-// CHECK13-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK13:       omp_offload.cont21:
+// CHECK13-NEXT:    [[TMP76:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    store i32 [[TMP76]], i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK13-NEXT:    [[TMP77:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK13-NEXT:    [[CONV24:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED23]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP77]], i32* [[CONV24]], align 4
+// CHECK13-NEXT:    [[TMP78:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED23]], align 8
+// CHECK13-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP80]], align 8
+// CHECK13-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP82]], align 8
+// CHECK13-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP83]], align 8
+// CHECK13-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP78]], i64* [[TMP85]], align 8
+// CHECK13-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP78]], i64* [[TMP87]], align 8
+// CHECK13-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP88]], align 8
+// CHECK13-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, i32 2, i8** [[TMP57]], i8** [[TMP58]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK13-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED26:%.*]], label [[OMP_OFFLOAD_CONT27:%.*]]
-// CHECK13:       omp_offload.failed26:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i64 [[TMP46]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT27]]
-// CHECK13:       omp_offload.cont27:
+// CHECK13-NEXT:    [[KERNEL_ARGS29:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP91]], align 4
+// CHECK13-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 2, i32* [[TMP92]], align 4
+// CHECK13-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP89]], i8*** [[TMP93]], align 8
+// CHECK13-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP90]], i8*** [[TMP94]], align 8
+// CHECK13-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP95]], align 8
+// CHECK13-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP96]], align 8
+// CHECK13-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP97]], align 8
+// CHECK13-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP98]], align 8
+// CHECK13-NEXT:    [[TMP99:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]])
+// CHECK13-NEXT:    [[TMP100:%.*]] = icmp ne i32 [[TMP99]], 0
+// CHECK13-NEXT:    br i1 [[TMP100]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
+// CHECK13:       omp_offload.failed30:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i64 [[TMP78]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
+// CHECK13:       omp_offload.cont31:
 // CHECK13-NEXT:    ret i32 0
 //
 //
@@ -7182,34 +7692,34 @@ int main (int argc, char **argv) {
 // CHECK15-NEXT:    [[_TMP8:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_17:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[N_CASTED19:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
-// CHECK15-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES23:%.*]] = alloca [4 x i64], align 4
+// CHECK15-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[N_CASTED32:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [3 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [3 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [3 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES36:%.*]] = alloca [3 x i64], align 4
-// CHECK15-NEXT:    [[_TMP37:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_38:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_39:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_46:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[N_CASTED47:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR__CASTED48:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS49:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS50:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS51:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES52:%.*]] = alloca [4 x i64], align 4
-// CHECK15-NEXT:    [[_TMP53:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_54:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_55:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[N_CASTED34:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS35:%.*]] = alloca [3 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS36:%.*]] = alloca [3 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS37:%.*]] = alloca [3 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES38:%.*]] = alloca [3 x i64], align 4
+// CHECK15-NEXT:    [[_TMP39:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_40:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_41:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_49:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[N_CASTED50:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR__CASTED51:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS52:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS53:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS54:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES55:%.*]] = alloca [4 x i64], align 4
+// CHECK15-NEXT:    [[_TMP56:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_57:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_58:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK15-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK15-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -7267,262 +7777,347 @@ int main (int argc, char **argv) {
 // CHECK15-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK15-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK15-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK15-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK15-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK15-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK15-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK15-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK15-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK15-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK15-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK15-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK15-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK15-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK15-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK15-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK15:       omp_offload.failed:
 // CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK15:       omp_offload.cont:
-// CHECK15-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK15-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK15-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK15-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK15-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
-// CHECK15-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK15-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK15-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK15-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK15-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK15-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK15-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK15-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK15-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK15-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK15-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK15-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK15-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK15-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK15-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK15-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK15-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK15-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK15-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK15-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
+// CHECK15-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK15-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK15-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK15-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK15-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK15-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK15-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK15-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK15-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK15-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK15-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK15-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK15-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK15-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK15-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK15-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK15-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK15-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK15-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK15-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK15-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP59]])
-// CHECK15-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK15-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK15:       omp_offload.failed15:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK15:       omp_offload.cont16:
-// CHECK15-NEXT:    [[TMP62:%.*]] = load i32, i32* [[M]], align 4
-// CHECK15-NEXT:    store i32 [[TMP62]], i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK15-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP63]], i32* [[N_CASTED18]], align 4
-// CHECK15-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N_CASTED18]], align 4
-// CHECK15-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK15-NEXT:    store i32 [[TMP65]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK15-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK15-NEXT:    [[TMP67:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK15-NEXT:    [[TMP68:%.*]] = sext i32 [[TMP67]] to i64
-// CHECK15-NEXT:    [[TMP69:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
-// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP69]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
-// CHECK15-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP64]], i32* [[TMP71]], align 4
-// CHECK15-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP64]], i32* [[TMP73]], align 4
-// CHECK15-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP74]], align 4
-// CHECK15-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP76]], align 4
-// CHECK15-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP78]], align 4
-// CHECK15-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP79]], align 4
-// CHECK15-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 4
-// CHECK15-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 4
-// CHECK15-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 2
-// CHECK15-NEXT:    store i64 [[TMP68]], i64* [[TMP84]], align 4
-// CHECK15-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
-// CHECK15-NEXT:    store i8* null, i8** [[TMP85]], align 4
-// CHECK15-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK15-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK15-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP67]])
+// CHECK15-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK15-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK15-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK15-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK15-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK15-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK15-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK15-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK15-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK15-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK15-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK15:       omp_offload.failed16:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK15:       omp_offload.cont17:
+// CHECK15-NEXT:    [[TMP78:%.*]] = load i32, i32* [[M]], align 4
+// CHECK15-NEXT:    store i32 [[TMP78]], i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK15-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP79]], i32* [[N_CASTED19]], align 4
+// CHECK15-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_CASTED19]], align 4
+// CHECK15-NEXT:    [[TMP81:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK15-NEXT:    store i32 [[TMP81]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK15-NEXT:    [[TMP82:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK15-NEXT:    [[TMP83:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK15-NEXT:    [[TMP84:%.*]] = sext i32 [[TMP83]] to i64
+// CHECK15-NEXT:    [[TMP85:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES23]] to i8*
+// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP85]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
+// CHECK15-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
 // CHECK15-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP66]], i32* [[TMP87]], align 4
-// CHECK15-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK15-NEXT:    store i32 [[TMP80]], i32* [[TMP87]], align 4
+// CHECK15-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
 // CHECK15-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP66]], i32* [[TMP89]], align 4
-// CHECK15-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK15-NEXT:    store i32 [[TMP80]], i32* [[TMP89]], align 4
+// CHECK15-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
 // CHECK15-NEXT:    store i8* null, i8** [[TMP90]], align 4
-// CHECK15-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP94:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP94]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK15-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK15-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP95]], 0
-// CHECK15-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
-// CHECK15-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
-// CHECK15-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK15-NEXT:    [[TMP96:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK15-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP96]], 1
-// CHECK15-NEXT:    [[TMP97:%.*]] = zext i32 [[ADD29]] to i64
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP97]])
-// CHECK15-NEXT:    [[TMP98:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, i32 4, i8** [[TMP91]], i8** [[TMP92]], i64* [[TMP93]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP99:%.*]] = icmp ne i32 [[TMP98]], 0
-// CHECK15-NEXT:    br i1 [[TMP99]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK15:       omp_offload.failed30:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i32 [[TMP64]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP66]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK15:       omp_offload.cont31:
-// CHECK15-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP100]], i32* [[N_CASTED32]], align 4
-// CHECK15-NEXT:    [[TMP101:%.*]] = load i32, i32* [[N_CASTED32]], align 4
-// CHECK15-NEXT:    [[TMP102:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK15-NEXT:    [[TMP103:%.*]] = sext i32 [[TMP102]] to i64
-// CHECK15-NEXT:    [[TMP104:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES36]] to i8*
-// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP104]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
-// CHECK15-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP106:%.*]] = bitcast i8** [[TMP105]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP101]], i32* [[TMP106]], align 4
-// CHECK15-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP101]], i32* [[TMP108]], align 4
-// CHECK15-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP109]], align 4
-// CHECK15-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP111:%.*]] = bitcast i8** [[TMP110]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP111]], align 4
-// CHECK15-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP113]], align 4
-// CHECK15-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP114]], align 4
-// CHECK15-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP116]], align 4
-// CHECK15-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP118]], align 4
-// CHECK15-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES36]], i32 0, i32 2
-// CHECK15-NEXT:    store i64 [[TMP103]], i64* [[TMP119]], align 4
-// CHECK15-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 2
-// CHECK15-NEXT:    store i8* null, i8** [[TMP120]], align 4
-// CHECK15-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES36]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP92]], align 4
+// CHECK15-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP94]], align 4
+// CHECK15-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP95]], align 4
+// CHECK15-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 4
+// CHECK15-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 4
+// CHECK15-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 2
+// CHECK15-NEXT:    store i64 [[TMP84]], i64* [[TMP100]], align 4
+// CHECK15-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 2
+// CHECK15-NEXT:    store i8* null, i8** [[TMP101]], align 4
+// CHECK15-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP82]], i32* [[TMP103]], align 4
+// CHECK15-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP82]], i32* [[TMP105]], align 4
+// CHECK15-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 3
+// CHECK15-NEXT:    store i8* null, i8** [[TMP106]], align 4
+// CHECK15-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP110:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP110]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK15-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK15-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP111]], 0
+// CHECK15-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
+// CHECK15-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
+// CHECK15-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK15-NEXT:    [[TMP112:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK15-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP112]], 1
+// CHECK15-NEXT:    [[TMP113:%.*]] = zext i32 [[ADD30]] to i64
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP113]])
+// CHECK15-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP114]], align 4
+// CHECK15-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 4, i32* [[TMP115]], align 4
+// CHECK15-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK15-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP108]], i8*** [[TMP117]], align 4
+// CHECK15-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP109]], i64** [[TMP118]], align 4
+// CHECK15-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP119]], align 4
+// CHECK15-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK15-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP121]], align 4
+// CHECK15-NEXT:    [[TMP122:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK15-NEXT:    [[TMP123:%.*]] = icmp ne i32 [[TMP122]], 0
+// CHECK15-NEXT:    br i1 [[TMP123]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK15:       omp_offload.failed32:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i32 [[TMP80]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP82]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK15:       omp_offload.cont33:
 // CHECK15-NEXT:    [[TMP124:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP124]], i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK15-NEXT:    [[TMP125:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK15-NEXT:    [[SUB40:%.*]] = sub nsw i32 [[TMP125]], 0
-// CHECK15-NEXT:    [[DIV41:%.*]] = sdiv i32 [[SUB40]], 1
-// CHECK15-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[DIV41]], 1
-// CHECK15-NEXT:    store i32 [[SUB42]], i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK15-NEXT:    [[TMP126:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK15-NEXT:    [[ADD43:%.*]] = add nsw i32 [[TMP126]], 1
-// CHECK15-NEXT:    [[TMP127:%.*]] = zext i32 [[ADD43]] to i64
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP127]])
-// CHECK15-NEXT:    [[TMP128:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, i32 3, i8** [[TMP121]], i8** [[TMP122]], i64* [[TMP123]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP129:%.*]] = icmp ne i32 [[TMP128]], 0
-// CHECK15-NEXT:    br i1 [[TMP129]], label [[OMP_OFFLOAD_FAILED44:%.*]], label [[OMP_OFFLOAD_CONT45:%.*]]
-// CHECK15:       omp_offload.failed44:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i32 [[TMP101]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT45]]
-// CHECK15:       omp_offload.cont45:
-// CHECK15-NEXT:    [[TMP130:%.*]] = load i32, i32* [[M]], align 4
-// CHECK15-NEXT:    store i32 [[TMP130]], i32* [[DOTCAPTURE_EXPR_46]], align 4
-// CHECK15-NEXT:    [[TMP131:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP131]], i32* [[N_CASTED47]], align 4
-// CHECK15-NEXT:    [[TMP132:%.*]] = load i32, i32* [[N_CASTED47]], align 4
-// CHECK15-NEXT:    [[TMP133:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_46]], align 4
-// CHECK15-NEXT:    store i32 [[TMP133]], i32* [[DOTCAPTURE_EXPR__CASTED48]], align 4
-// CHECK15-NEXT:    [[TMP134:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED48]], align 4
-// CHECK15-NEXT:    [[TMP135:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK15-NEXT:    [[TMP136:%.*]] = sext i32 [[TMP135]] to i64
-// CHECK15-NEXT:    [[TMP137:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES52]] to i8*
-// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP137]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
-// CHECK15-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP132]], i32* [[TMP139]], align 4
-// CHECK15-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP132]], i32* [[TMP141]], align 4
-// CHECK15-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP142]], align 4
-// CHECK15-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP144]], align 4
-// CHECK15-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP146]], align 4
-// CHECK15-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP147]], align 4
-// CHECK15-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP149]], align 4
-// CHECK15-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP151]], align 4
-// CHECK15-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES52]], i32 0, i32 2
-// CHECK15-NEXT:    store i64 [[TMP136]], i64* [[TMP152]], align 4
-// CHECK15-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 2
-// CHECK15-NEXT:    store i8* null, i8** [[TMP153]], align 4
-// CHECK15-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 3
-// CHECK15-NEXT:    [[TMP155:%.*]] = bitcast i8** [[TMP154]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP134]], i32* [[TMP155]], align 4
-// CHECK15-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 3
-// CHECK15-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP134]], i32* [[TMP157]], align 4
-// CHECK15-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 3
-// CHECK15-NEXT:    store i8* null, i8** [[TMP158]], align 4
-// CHECK15-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES52]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP162:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP162]], i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK15-NEXT:    [[TMP163:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK15-NEXT:    [[SUB56:%.*]] = sub nsw i32 [[TMP163]], 0
-// CHECK15-NEXT:    [[DIV57:%.*]] = sdiv i32 [[SUB56]], 1
-// CHECK15-NEXT:    [[SUB58:%.*]] = sub nsw i32 [[DIV57]], 1
-// CHECK15-NEXT:    store i32 [[SUB58]], i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK15-NEXT:    [[TMP164:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK15-NEXT:    [[ADD59:%.*]] = add nsw i32 [[TMP164]], 1
-// CHECK15-NEXT:    [[TMP165:%.*]] = zext i32 [[ADD59]] to i64
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP165]])
-// CHECK15-NEXT:    [[TMP166:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, i32 4, i8** [[TMP159]], i8** [[TMP160]], i64* [[TMP161]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP167:%.*]] = icmp ne i32 [[TMP166]], 0
-// CHECK15-NEXT:    br i1 [[TMP167]], label [[OMP_OFFLOAD_FAILED60:%.*]], label [[OMP_OFFLOAD_CONT61:%.*]]
-// CHECK15:       omp_offload.failed60:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i32 [[TMP132]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP134]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT61]]
-// CHECK15:       omp_offload.cont61:
-// CHECK15-NEXT:    [[TMP168:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK15-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP168]])
+// CHECK15-NEXT:    store i32 [[TMP124]], i32* [[N_CASTED34]], align 4
+// CHECK15-NEXT:    [[TMP125:%.*]] = load i32, i32* [[N_CASTED34]], align 4
+// CHECK15-NEXT:    [[TMP126:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK15-NEXT:    [[TMP127:%.*]] = sext i32 [[TMP126]] to i64
+// CHECK15-NEXT:    [[TMP128:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES38]] to i8*
+// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP128]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
+// CHECK15-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP125]], i32* [[TMP130]], align 4
+// CHECK15-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP125]], i32* [[TMP132]], align 4
+// CHECK15-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS37]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP133]], align 4
+// CHECK15-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP135]], align 4
+// CHECK15-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP137]], align 4
+// CHECK15-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS37]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP138]], align 4
+// CHECK15-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP140]], align 4
+// CHECK15-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP142]], align 4
+// CHECK15-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES38]], i32 0, i32 2
+// CHECK15-NEXT:    store i64 [[TMP127]], i64* [[TMP143]], align 4
+// CHECK15-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS37]], i32 0, i32 2
+// CHECK15-NEXT:    store i8* null, i8** [[TMP144]], align 4
+// CHECK15-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES38]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP148:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP148]], i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK15-NEXT:    [[TMP149:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK15-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[TMP149]], 0
+// CHECK15-NEXT:    [[DIV43:%.*]] = sdiv i32 [[SUB42]], 1
+// CHECK15-NEXT:    [[SUB44:%.*]] = sub nsw i32 [[DIV43]], 1
+// CHECK15-NEXT:    store i32 [[SUB44]], i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK15-NEXT:    [[TMP150:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK15-NEXT:    [[ADD45:%.*]] = add nsw i32 [[TMP150]], 1
+// CHECK15-NEXT:    [[TMP151:%.*]] = zext i32 [[ADD45]] to i64
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP151]])
+// CHECK15-NEXT:    [[KERNEL_ARGS46:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP152]], align 4
+// CHECK15-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 3, i32* [[TMP153]], align 4
+// CHECK15-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP145]], i8*** [[TMP154]], align 4
+// CHECK15-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP146]], i8*** [[TMP155]], align 4
+// CHECK15-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP147]], i64** [[TMP156]], align 4
+// CHECK15-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP157]], align 4
+// CHECK15-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP158]], align 4
+// CHECK15-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP159]], align 4
+// CHECK15-NEXT:    [[TMP160:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]])
+// CHECK15-NEXT:    [[TMP161:%.*]] = icmp ne i32 [[TMP160]], 0
+// CHECK15-NEXT:    br i1 [[TMP161]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
+// CHECK15:       omp_offload.failed47:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i32 [[TMP125]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
+// CHECK15:       omp_offload.cont48:
+// CHECK15-NEXT:    [[TMP162:%.*]] = load i32, i32* [[M]], align 4
+// CHECK15-NEXT:    store i32 [[TMP162]], i32* [[DOTCAPTURE_EXPR_49]], align 4
+// CHECK15-NEXT:    [[TMP163:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP163]], i32* [[N_CASTED50]], align 4
+// CHECK15-NEXT:    [[TMP164:%.*]] = load i32, i32* [[N_CASTED50]], align 4
+// CHECK15-NEXT:    [[TMP165:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_49]], align 4
+// CHECK15-NEXT:    store i32 [[TMP165]], i32* [[DOTCAPTURE_EXPR__CASTED51]], align 4
+// CHECK15-NEXT:    [[TMP166:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED51]], align 4
+// CHECK15-NEXT:    [[TMP167:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK15-NEXT:    [[TMP168:%.*]] = sext i32 [[TMP167]] to i64
+// CHECK15-NEXT:    [[TMP169:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES55]] to i8*
+// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP169]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
+// CHECK15-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP171:%.*]] = bitcast i8** [[TMP170]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP164]], i32* [[TMP171]], align 4
+// CHECK15-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP173:%.*]] = bitcast i8** [[TMP172]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP164]], i32* [[TMP173]], align 4
+// CHECK15-NEXT:    [[TMP174:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP174]], align 4
+// CHECK15-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP176:%.*]] = bitcast i8** [[TMP175]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP176]], align 4
+// CHECK15-NEXT:    [[TMP177:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP178:%.*]] = bitcast i8** [[TMP177]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP178]], align 4
+// CHECK15-NEXT:    [[TMP179:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP179]], align 4
+// CHECK15-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP181]], align 4
+// CHECK15-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP183:%.*]] = bitcast i8** [[TMP182]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP183]], align 4
+// CHECK15-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES55]], i32 0, i32 2
+// CHECK15-NEXT:    store i64 [[TMP168]], i64* [[TMP184]], align 4
+// CHECK15-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 2
+// CHECK15-NEXT:    store i8* null, i8** [[TMP185]], align 4
+// CHECK15-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP187:%.*]] = bitcast i8** [[TMP186]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP166]], i32* [[TMP187]], align 4
+// CHECK15-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP189:%.*]] = bitcast i8** [[TMP188]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP166]], i32* [[TMP189]], align 4
+// CHECK15-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 3
+// CHECK15-NEXT:    store i8* null, i8** [[TMP190]], align 4
+// CHECK15-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES55]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP194:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP194]], i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK15-NEXT:    [[TMP195:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK15-NEXT:    [[SUB59:%.*]] = sub nsw i32 [[TMP195]], 0
+// CHECK15-NEXT:    [[DIV60:%.*]] = sdiv i32 [[SUB59]], 1
+// CHECK15-NEXT:    [[SUB61:%.*]] = sub nsw i32 [[DIV60]], 1
+// CHECK15-NEXT:    store i32 [[SUB61]], i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK15-NEXT:    [[TMP196:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK15-NEXT:    [[ADD62:%.*]] = add nsw i32 [[TMP196]], 1
+// CHECK15-NEXT:    [[TMP197:%.*]] = zext i32 [[ADD62]] to i64
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP197]])
+// CHECK15-NEXT:    [[KERNEL_ARGS63:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP198]], align 4
+// CHECK15-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 4, i32* [[TMP199]], align 4
+// CHECK15-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP191]], i8*** [[TMP200]], align 4
+// CHECK15-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP192]], i8*** [[TMP201]], align 4
+// CHECK15-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP193]], i64** [[TMP202]], align 4
+// CHECK15-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP203]], align 4
+// CHECK15-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP204]], align 4
+// CHECK15-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP205]], align 4
+// CHECK15-NEXT:    [[TMP206:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]])
+// CHECK15-NEXT:    [[TMP207:%.*]] = icmp ne i32 [[TMP206]], 0
+// CHECK15-NEXT:    br i1 [[TMP207]], label [[OMP_OFFLOAD_FAILED64:%.*]], label [[OMP_OFFLOAD_CONT65:%.*]]
+// CHECK15:       omp_offload.failed64:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i32 [[TMP164]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP166]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT65]]
+// CHECK15:       omp_offload.cont65:
+// CHECK15-NEXT:    [[TMP208:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK15-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP208]])
 // CHECK15-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK15-NEXT:    [[TMP169:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK15-NEXT:    call void @llvm.stackrestore(i8* [[TMP169]])
-// CHECK15-NEXT:    [[TMP170:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK15-NEXT:    ret i32 [[TMP170]]
+// CHECK15-NEXT:    [[TMP209:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK15-NEXT:    call void @llvm.stackrestore(i8* [[TMP209]])
+// CHECK15-NEXT:    [[TMP210:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK15-NEXT:    ret i32 [[TMP210]]
 //
 //
 // CHECK15-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139
@@ -8767,20 +9362,20 @@ int main (int argc, char **argv) {
 // CHECK15-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 4
-// CHECK15-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR__CASTED20:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 4
+// CHECK15-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR__CASTED23:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK15-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK15-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -8794,111 +9389,196 @@ int main (int argc, char **argv) {
 // CHECK15-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK15-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK15-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK15-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK15-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK15-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK15-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK15-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK15-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK15-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK15-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK15-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK15-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK15-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK15-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK15:       omp_offload.failed:
 // CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK15:       omp_offload.cont:
-// CHECK15-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK15-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK15-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK15-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK15-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK15-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK15-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK15-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK15:       omp_offload.failed5:
+// CHECK15-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK15-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK15-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK15-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK15-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK15-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK15-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK15-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK15-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK15-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK15-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK15:       omp_offload.failed6:
 // CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK15:       omp_offload.cont6:
-// CHECK15-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK15-NEXT:    store i32 [[TMP18]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK15-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK15-NEXT:    store i32 [[TMP19]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK15-NEXT:    [[TMP20:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK15-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP22]], align 4
-// CHECK15-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP24]], align 4
-// CHECK15-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP25]], align 4
-// CHECK15-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP20]], i32* [[TMP27]], align 4
-// CHECK15-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP20]], i32* [[TMP29]], align 4
-// CHECK15-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP30]], align 4
-// CHECK15-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, i32 2, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK15-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK15:       omp_offload.failed11:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i32 [[TMP20]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK15:       omp_offload.cont12:
-// CHECK15-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP36]], align 4
-// CHECK15-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK15:       omp_offload.cont7:
+// CHECK15-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK15-NEXT:    store i32 [[TMP34]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK15-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK15-NEXT:    store i32 [[TMP35]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK15-NEXT:    [[TMP36:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK15-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
 // CHECK15-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [10 x i32]**
 // CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP38]], align 4
-// CHECK15-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP39]], align 4
-// CHECK15-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP40]], align 4
+// CHECK15-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP41]], align 4
+// CHECK15-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP36]], i32* [[TMP43]], align 4
+// CHECK15-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP36]], i32* [[TMP45]], align 4
+// CHECK15-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP46]], align 4
+// CHECK15-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
+// CHECK15-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP49]], align 4
+// CHECK15-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 2, i32* [[TMP50]], align 4
+// CHECK15-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 4
+// CHECK15-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP48]], i8*** [[TMP52]], align 4
+// CHECK15-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP53]], align 4
+// CHECK15-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP54]], align 4
+// CHECK15-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP55]], align 4
+// CHECK15-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP56]], align 4
+// CHECK15-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK15-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
+// CHECK15-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK15:       omp_offload.failed13:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i32 [[TMP36]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK15:       omp_offload.cont14:
+// CHECK15-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP60]], align 4
+// CHECK15-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP62]], align 4
+// CHECK15-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP63]], align 4
+// CHECK15-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, i32 1, i8** [[TMP40]], i8** [[TMP41]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK15-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK15:       omp_offload.failed17:
+// CHECK15-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK15-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK15-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK15-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 4
+// CHECK15-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK15-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP71]], align 4
+// CHECK15-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK15-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP73]], align 4
+// CHECK15-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK15-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK15-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK15:       omp_offload.failed20:
 // CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK15:       omp_offload.cont18:
-// CHECK15-NEXT:    [[TMP44:%.*]] = load i32, i32* [[M]], align 4
-// CHECK15-NEXT:    store i32 [[TMP44]], i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK15-NEXT:    [[TMP45:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK15-NEXT:    store i32 [[TMP45]], i32* [[DOTCAPTURE_EXPR__CASTED20]], align 4
-// CHECK15-NEXT:    [[TMP46:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED20]], align 4
-// CHECK15-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP48]], align 4
-// CHECK15-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP50]], align 4
-// CHECK15-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP51]], align 4
-// CHECK15-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP46]], i32* [[TMP53]], align 4
-// CHECK15-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP46]], i32* [[TMP55]], align 4
-// CHECK15-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP56]], align 4
-// CHECK15-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK15:       omp_offload.cont21:
+// CHECK15-NEXT:    [[TMP76:%.*]] = load i32, i32* [[M]], align 4
+// CHECK15-NEXT:    store i32 [[TMP76]], i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK15-NEXT:    [[TMP77:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK15-NEXT:    store i32 [[TMP77]], i32* [[DOTCAPTURE_EXPR__CASTED23]], align 4
+// CHECK15-NEXT:    [[TMP78:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED23]], align 4
+// CHECK15-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP80]], align 4
+// CHECK15-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP82]], align 4
+// CHECK15-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP83]], align 4
+// CHECK15-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP78]], i32* [[TMP85]], align 4
+// CHECK15-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP78]], i32* [[TMP87]], align 4
+// CHECK15-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP88]], align 4
+// CHECK15-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, i32 2, i8** [[TMP57]], i8** [[TMP58]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK15-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
-// CHECK15:       omp_offload.failed25:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i32 [[TMP46]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
-// CHECK15:       omp_offload.cont26:
+// CHECK15-NEXT:    [[KERNEL_ARGS28:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP91]], align 4
+// CHECK15-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 2, i32* [[TMP92]], align 4
+// CHECK15-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP89]], i8*** [[TMP93]], align 4
+// CHECK15-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP90]], i8*** [[TMP94]], align 4
+// CHECK15-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP95]], align 4
+// CHECK15-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP96]], align 4
+// CHECK15-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP97]], align 4
+// CHECK15-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP98]], align 4
+// CHECK15-NEXT:    [[TMP99:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]])
+// CHECK15-NEXT:    [[TMP100:%.*]] = icmp ne i32 [[TMP99]], 0
+// CHECK15-NEXT:    br i1 [[TMP100]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
+// CHECK15:       omp_offload.failed29:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i32 [[TMP78]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
+// CHECK15:       omp_offload.cont30:
 // CHECK15-NEXT:    ret i32 0
 //
 //
@@ -9775,34 +10455,34 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[_TMP9:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
-// CHECK17-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES26:%.*]] = alloca [4 x i64], align 8
+// CHECK17-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[N_CASTED35:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS37:%.*]] = alloca [3 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS38:%.*]] = alloca [3 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS39:%.*]] = alloca [3 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES40:%.*]] = alloca [3 x i64], align 8
-// CHECK17-NEXT:    [[_TMP41:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_42:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_43:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_50:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[N_CASTED51:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED53:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS55:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS56:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS57:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES58:%.*]] = alloca [4 x i64], align 8
-// CHECK17-NEXT:    [[_TMP59:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_60:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_61:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_29:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[N_CASTED37:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS39:%.*]] = alloca [3 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS40:%.*]] = alloca [3 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS41:%.*]] = alloca [3 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES42:%.*]] = alloca [3 x i64], align 8
+// CHECK17-NEXT:    [[_TMP43:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_44:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_45:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_53:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[N_CASTED54:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED56:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS58:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS59:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS60:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES61:%.*]] = alloca [4 x i64], align 8
+// CHECK17-NEXT:    [[_TMP62:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_63:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_64:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK17-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK17-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -9861,264 +10541,349 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK17-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK17-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK17-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK17-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK17-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK17-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK17-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK17-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK17-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK17-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
-// CHECK17-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK17-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK17-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK17-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK17-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
-// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK17-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK17-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK17-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK17-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK17-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK17-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK17-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK17-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK17-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK17-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK17-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK17-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK17-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
+// CHECK17-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK17-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK17-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK17-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK17-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK17-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK17-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK17-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK17-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK17-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK17-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK17-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK17-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK17-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK17-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK17-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK17-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP58]])
-// CHECK17-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK17-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK17:       omp_offload.failed16:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK17:       omp_offload.cont17:
-// CHECK17-NEXT:    [[TMP61:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    store i32 [[TMP61]], i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK17-NEXT:    [[TMP62:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP62]], i32* [[CONV20]], align 4
-// CHECK17-NEXT:    [[TMP63:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK17-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK17-NEXT:    [[CONV21:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP64]], i32* [[CONV21]], align 4
-// CHECK17-NEXT:    [[TMP65:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK17-NEXT:    [[TMP66:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK17-NEXT:    [[TMP67:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP67]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
-// CHECK17-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP63]], i64* [[TMP69]], align 8
-// CHECK17-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP63]], i64* [[TMP71]], align 8
-// CHECK17-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP72]], align 8
-// CHECK17-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP74]], align 8
-// CHECK17-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP76]], align 8
-// CHECK17-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP77]], align 8
-// CHECK17-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP79]], align 8
-// CHECK17-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 8
-// CHECK17-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 2
-// CHECK17-NEXT:    store i64 [[TMP66]], i64* [[TMP82]], align 8
-// CHECK17-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP83]], align 8
-// CHECK17-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK17-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK17-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP66]])
+// CHECK17-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK17-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK17-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK17-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK17-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK17-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK17-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK17-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK17-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK17-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK17-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK17:       omp_offload.failed17:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK17:       omp_offload.cont18:
+// CHECK17-NEXT:    [[TMP77:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    store i32 [[TMP77]], i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK17-NEXT:    [[TMP78:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP78]], i32* [[CONV21]], align 4
+// CHECK17-NEXT:    [[TMP79:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK17-NEXT:    [[TMP80:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK17-NEXT:    [[CONV22:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP80]], i32* [[CONV22]], align 4
+// CHECK17-NEXT:    [[TMP81:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK17-NEXT:    [[TMP82:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK17-NEXT:    [[TMP83:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES26]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP83]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
+// CHECK17-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP65]], i64* [[TMP85]], align 8
-// CHECK17-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK17-NEXT:    store i64 [[TMP79]], i64* [[TMP85]], align 8
+// CHECK17-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP65]], i64* [[TMP87]], align 8
-// CHECK17-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK17-NEXT:    store i64 [[TMP79]], i64* [[TMP87]], align 8
+// CHECK17-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 0
 // CHECK17-NEXT:    store i8* null, i8** [[TMP88]], align 8
-// CHECK17-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP92:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP92]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK17-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK17-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP93]], 0
-// CHECK17-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
-// CHECK17-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
-// CHECK17-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK17-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK17-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP94]], 1
-// CHECK17-NEXT:    [[TMP95:%.*]] = zext i32 [[ADD32]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP95]])
-// CHECK17-NEXT:    [[TMP96:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, i32 4, i8** [[TMP89]], i8** [[TMP90]], i64* [[TMP91]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP97:%.*]] = icmp ne i32 [[TMP96]], 0
-// CHECK17-NEXT:    br i1 [[TMP97]], label [[OMP_OFFLOAD_FAILED33:%.*]], label [[OMP_OFFLOAD_CONT34:%.*]]
-// CHECK17:       omp_offload.failed33:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i64 [[TMP63]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP65]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT34]]
-// CHECK17:       omp_offload.cont34:
-// CHECK17-NEXT:    [[TMP98:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    [[CONV36:%.*]] = bitcast i64* [[N_CASTED35]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP98]], i32* [[CONV36]], align 4
-// CHECK17-NEXT:    [[TMP99:%.*]] = load i64, i64* [[N_CASTED35]], align 8
-// CHECK17-NEXT:    [[TMP100:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK17-NEXT:    [[TMP101:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES40]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP101]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
-// CHECK17-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP90]], align 8
+// CHECK17-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP92]], align 8
+// CHECK17-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP93]], align 8
+// CHECK17-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP95]], align 8
+// CHECK17-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 8
+// CHECK17-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 2
+// CHECK17-NEXT:    store i64 [[TMP82]], i64* [[TMP98]], align 8
+// CHECK17-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP99]], align 8
+// CHECK17-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP81]], i64* [[TMP101]], align 8
+// CHECK17-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 3
 // CHECK17-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP99]], i64* [[TMP103]], align 8
-// CHECK17-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP99]], i64* [[TMP105]], align 8
-// CHECK17-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS39]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP106]], align 8
-// CHECK17-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP108]], align 8
-// CHECK17-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP110]], align 8
-// CHECK17-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS39]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP111]], align 8
-// CHECK17-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP113]], align 8
-// CHECK17-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP115]], align 8
-// CHECK17-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES40]], i32 0, i32 2
-// CHECK17-NEXT:    store i64 [[TMP100]], i64* [[TMP116]], align 8
-// CHECK17-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS39]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP117]], align 8
-// CHECK17-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS37]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS38]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES40]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP121:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP121]], i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK17-NEXT:    [[TMP122:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK17-NEXT:    [[SUB44:%.*]] = sub nsw i32 [[TMP122]], 0
-// CHECK17-NEXT:    [[DIV45:%.*]] = sdiv i32 [[SUB44]], 1
-// CHECK17-NEXT:    [[SUB46:%.*]] = sub nsw i32 [[DIV45]], 1
-// CHECK17-NEXT:    store i32 [[SUB46]], i32* [[DOTCAPTURE_EXPR_43]], align 4
-// CHECK17-NEXT:    [[TMP123:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_43]], align 4
-// CHECK17-NEXT:    [[ADD47:%.*]] = add nsw i32 [[TMP123]], 1
-// CHECK17-NEXT:    [[TMP124:%.*]] = zext i32 [[ADD47]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP124]])
-// CHECK17-NEXT:    [[TMP125:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, i32 3, i8** [[TMP118]], i8** [[TMP119]], i64* [[TMP120]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP126:%.*]] = icmp ne i32 [[TMP125]], 0
-// CHECK17-NEXT:    br i1 [[TMP126]], label [[OMP_OFFLOAD_FAILED48:%.*]], label [[OMP_OFFLOAD_CONT49:%.*]]
-// CHECK17:       omp_offload.failed48:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i64 [[TMP99]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT49]]
-// CHECK17:       omp_offload.cont49:
-// CHECK17-NEXT:    [[TMP127:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    store i32 [[TMP127]], i32* [[DOTCAPTURE_EXPR_50]], align 4
-// CHECK17-NEXT:    [[TMP128:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    [[CONV52:%.*]] = bitcast i64* [[N_CASTED51]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP128]], i32* [[CONV52]], align 4
-// CHECK17-NEXT:    [[TMP129:%.*]] = load i64, i64* [[N_CASTED51]], align 8
-// CHECK17-NEXT:    [[TMP130:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_50]], align 4
-// CHECK17-NEXT:    [[CONV54:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED53]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP130]], i32* [[CONV54]], align 4
-// CHECK17-NEXT:    [[TMP131:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED53]], align 8
-// CHECK17-NEXT:    [[TMP132:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK17-NEXT:    [[TMP133:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES58]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP133]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
-// CHECK17-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP129]], i64* [[TMP135]], align 8
-// CHECK17-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP129]], i64* [[TMP137]], align 8
-// CHECK17-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP138]], align 8
-// CHECK17-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP140]], align 8
-// CHECK17-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP142]], align 8
-// CHECK17-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP143]], align 8
-// CHECK17-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP145:%.*]] = bitcast i8** [[TMP144]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP145]], align 8
-// CHECK17-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP147]], align 8
-// CHECK17-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES58]], i32 0, i32 2
-// CHECK17-NEXT:    store i64 [[TMP132]], i64* [[TMP148]], align 8
-// CHECK17-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP149]], align 8
-// CHECK17-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP131]], i64* [[TMP151]], align 8
-// CHECK17-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP131]], i64* [[TMP153]], align 8
-// CHECK17-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS57]], i64 0, i64 3
-// CHECK17-NEXT:    store i8* null, i8** [[TMP154]], align 8
-// CHECK17-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS55]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS56]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES58]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP158:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP158]], i32* [[DOTCAPTURE_EXPR_60]], align 4
-// CHECK17-NEXT:    [[TMP159:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_60]], align 4
-// CHECK17-NEXT:    [[SUB62:%.*]] = sub nsw i32 [[TMP159]], 0
-// CHECK17-NEXT:    [[DIV63:%.*]] = sdiv i32 [[SUB62]], 1
-// CHECK17-NEXT:    [[SUB64:%.*]] = sub nsw i32 [[DIV63]], 1
-// CHECK17-NEXT:    store i32 [[SUB64]], i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK17-NEXT:    [[TMP160:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_61]], align 4
-// CHECK17-NEXT:    [[ADD65:%.*]] = add nsw i32 [[TMP160]], 1
-// CHECK17-NEXT:    [[TMP161:%.*]] = zext i32 [[ADD65]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP161]])
-// CHECK17-NEXT:    [[TMP162:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, i32 4, i8** [[TMP155]], i8** [[TMP156]], i64* [[TMP157]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP163:%.*]] = icmp ne i32 [[TMP162]], 0
-// CHECK17-NEXT:    br i1 [[TMP163]], label [[OMP_OFFLOAD_FAILED66:%.*]], label [[OMP_OFFLOAD_CONT67:%.*]]
-// CHECK17:       omp_offload.failed66:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i64 [[TMP129]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP131]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT67]]
-// CHECK17:       omp_offload.cont67:
-// CHECK17-NEXT:    [[TMP164:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK17-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP164]])
+// CHECK17-NEXT:    store i64 [[TMP81]], i64* [[TMP103]], align 8
+// CHECK17-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 3
+// CHECK17-NEXT:    store i8* null, i8** [[TMP104]], align 8
+// CHECK17-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP108:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP108]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK17-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK17-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[TMP109]], 0
+// CHECK17-NEXT:    [[DIV31:%.*]] = sdiv i32 [[SUB30]], 1
+// CHECK17-NEXT:    [[SUB32:%.*]] = sub nsw i32 [[DIV31]], 1
+// CHECK17-NEXT:    store i32 [[SUB32]], i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK17-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK17-NEXT:    [[ADD33:%.*]] = add nsw i32 [[TMP110]], 1
+// CHECK17-NEXT:    [[TMP111:%.*]] = zext i32 [[ADD33]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP111]])
+// CHECK17-NEXT:    [[KERNEL_ARGS34:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP112]], align 4
+// CHECK17-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 4, i32* [[TMP113]], align 4
+// CHECK17-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK17-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 8
+// CHECK17-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP107]], i64** [[TMP116]], align 8
+// CHECK17-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP117]], align 8
+// CHECK17-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK17-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP119]], align 8
+// CHECK17-NEXT:    [[TMP120:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]])
+// CHECK17-NEXT:    [[TMP121:%.*]] = icmp ne i32 [[TMP120]], 0
+// CHECK17-NEXT:    br i1 [[TMP121]], label [[OMP_OFFLOAD_FAILED35:%.*]], label [[OMP_OFFLOAD_CONT36:%.*]]
+// CHECK17:       omp_offload.failed35:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i64 [[TMP79]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP81]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT36]]
+// CHECK17:       omp_offload.cont36:
+// CHECK17-NEXT:    [[TMP122:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    [[CONV38:%.*]] = bitcast i64* [[N_CASTED37]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP122]], i32* [[CONV38]], align 4
+// CHECK17-NEXT:    [[TMP123:%.*]] = load i64, i64* [[N_CASTED37]], align 8
+// CHECK17-NEXT:    [[TMP124:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK17-NEXT:    [[TMP125:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES42]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP125]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
+// CHECK17-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP127:%.*]] = bitcast i8** [[TMP126]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP123]], i64* [[TMP127]], align 8
+// CHECK17-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP123]], i64* [[TMP129]], align 8
+// CHECK17-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS41]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP130]], align 8
+// CHECK17-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP132]], align 8
+// CHECK17-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP134]], align 8
+// CHECK17-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS41]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP135]], align 8
+// CHECK17-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP137]], align 8
+// CHECK17-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP139]], align 8
+// CHECK17-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES42]], i32 0, i32 2
+// CHECK17-NEXT:    store i64 [[TMP124]], i64* [[TMP140]], align 8
+// CHECK17-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS41]], i64 0, i64 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP141]], align 8
+// CHECK17-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS39]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS40]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES42]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP145:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP145]], i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK17-NEXT:    [[TMP146:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK17-NEXT:    [[SUB46:%.*]] = sub nsw i32 [[TMP146]], 0
+// CHECK17-NEXT:    [[DIV47:%.*]] = sdiv i32 [[SUB46]], 1
+// CHECK17-NEXT:    [[SUB48:%.*]] = sub nsw i32 [[DIV47]], 1
+// CHECK17-NEXT:    store i32 [[SUB48]], i32* [[DOTCAPTURE_EXPR_45]], align 4
+// CHECK17-NEXT:    [[TMP147:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_45]], align 4
+// CHECK17-NEXT:    [[ADD49:%.*]] = add nsw i32 [[TMP147]], 1
+// CHECK17-NEXT:    [[TMP148:%.*]] = zext i32 [[ADD49]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP148]])
+// CHECK17-NEXT:    [[KERNEL_ARGS50:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP149]], align 4
+// CHECK17-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP150]], align 4
+// CHECK17-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP142]], i8*** [[TMP151]], align 8
+// CHECK17-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP143]], i8*** [[TMP152]], align 8
+// CHECK17-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP144]], i64** [[TMP153]], align 8
+// CHECK17-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP154]], align 8
+// CHECK17-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP155]], align 8
+// CHECK17-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP156]], align 8
+// CHECK17-NEXT:    [[TMP157:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS50]])
+// CHECK17-NEXT:    [[TMP158:%.*]] = icmp ne i32 [[TMP157]], 0
+// CHECK17-NEXT:    br i1 [[TMP158]], label [[OMP_OFFLOAD_FAILED51:%.*]], label [[OMP_OFFLOAD_CONT52:%.*]]
+// CHECK17:       omp_offload.failed51:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i64 [[TMP123]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT52]]
+// CHECK17:       omp_offload.cont52:
+// CHECK17-NEXT:    [[TMP159:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    store i32 [[TMP159]], i32* [[DOTCAPTURE_EXPR_53]], align 4
+// CHECK17-NEXT:    [[TMP160:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    [[CONV55:%.*]] = bitcast i64* [[N_CASTED54]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP160]], i32* [[CONV55]], align 4
+// CHECK17-NEXT:    [[TMP161:%.*]] = load i64, i64* [[N_CASTED54]], align 8
+// CHECK17-NEXT:    [[TMP162:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_53]], align 4
+// CHECK17-NEXT:    [[CONV57:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED56]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP162]], i32* [[CONV57]], align 4
+// CHECK17-NEXT:    [[TMP163:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED56]], align 8
+// CHECK17-NEXT:    [[TMP164:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK17-NEXT:    [[TMP165:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES61]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP165]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
+// CHECK17-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP161]], i64* [[TMP167]], align 8
+// CHECK17-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP161]], i64* [[TMP169]], align 8
+// CHECK17-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP170]], align 8
+// CHECK17-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP172:%.*]] = bitcast i8** [[TMP171]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP172]], align 8
+// CHECK17-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP174:%.*]] = bitcast i8** [[TMP173]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP174]], align 8
+// CHECK17-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP175]], align 8
+// CHECK17-NEXT:    [[TMP176:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP177:%.*]] = bitcast i8** [[TMP176]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP177]], align 8
+// CHECK17-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP179]], align 8
+// CHECK17-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES61]], i32 0, i32 2
+// CHECK17-NEXT:    store i64 [[TMP164]], i64* [[TMP180]], align 8
+// CHECK17-NEXT:    [[TMP181:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP181]], align 8
+// CHECK17-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP183:%.*]] = bitcast i8** [[TMP182]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP163]], i64* [[TMP183]], align 8
+// CHECK17-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP185:%.*]] = bitcast i8** [[TMP184]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP163]], i64* [[TMP185]], align 8
+// CHECK17-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS60]], i64 0, i64 3
+// CHECK17-NEXT:    store i8* null, i8** [[TMP186]], align 8
+// CHECK17-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS58]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS59]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES61]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP190:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP190]], i32* [[DOTCAPTURE_EXPR_63]], align 4
+// CHECK17-NEXT:    [[TMP191:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_63]], align 4
+// CHECK17-NEXT:    [[SUB65:%.*]] = sub nsw i32 [[TMP191]], 0
+// CHECK17-NEXT:    [[DIV66:%.*]] = sdiv i32 [[SUB65]], 1
+// CHECK17-NEXT:    [[SUB67:%.*]] = sub nsw i32 [[DIV66]], 1
+// CHECK17-NEXT:    store i32 [[SUB67]], i32* [[DOTCAPTURE_EXPR_64]], align 4
+// CHECK17-NEXT:    [[TMP192:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_64]], align 4
+// CHECK17-NEXT:    [[ADD68:%.*]] = add nsw i32 [[TMP192]], 1
+// CHECK17-NEXT:    [[TMP193:%.*]] = zext i32 [[ADD68]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP193]])
+// CHECK17-NEXT:    [[KERNEL_ARGS69:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP194]], align 4
+// CHECK17-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 4, i32* [[TMP195]], align 4
+// CHECK17-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP187]], i8*** [[TMP196]], align 8
+// CHECK17-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP188]], i8*** [[TMP197]], align 8
+// CHECK17-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP189]], i64** [[TMP198]], align 8
+// CHECK17-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP199]], align 8
+// CHECK17-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP200]], align 8
+// CHECK17-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP201]], align 8
+// CHECK17-NEXT:    [[TMP202:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS69]])
+// CHECK17-NEXT:    [[TMP203:%.*]] = icmp ne i32 [[TMP202]], 0
+// CHECK17-NEXT:    br i1 [[TMP203]], label [[OMP_OFFLOAD_FAILED70:%.*]], label [[OMP_OFFLOAD_CONT71:%.*]]
+// CHECK17:       omp_offload.failed70:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i64 [[TMP161]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP163]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT71]]
+// CHECK17:       omp_offload.cont71:
+// CHECK17-NEXT:    [[TMP204:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK17-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP204]])
 // CHECK17-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK17-NEXT:    [[TMP165:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP165]])
-// CHECK17-NEXT:    [[TMP166:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK17-NEXT:    ret i32 [[TMP166]]
+// CHECK17-NEXT:    [[TMP205:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP205]])
+// CHECK17-NEXT:    [[TMP206:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK17-NEXT:    ret i32 [[TMP206]]
 //
 //
 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139
@@ -11423,20 +12188,20 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 8
-// CHECK17-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED20:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 8
+// CHECK17-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR__CASTED23:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS25:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS26:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS27:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[_TMP28:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK17-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK17-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -11450,113 +12215,198 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK17-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK17-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK17-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK17-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK17-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK17-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK17-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK17-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
-// CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK17-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK17-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK17-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK17-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK17-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK17:       omp_offload.failed5:
+// CHECK17-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK17-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK17-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK17-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK17-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK17-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK17:       omp_offload.failed6:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK17:       omp_offload.cont6:
-// CHECK17-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    store i32 [[TMP18]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK17-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK17:       omp_offload.cont7:
+// CHECK17-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    store i32 [[TMP34]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK17-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP19]], i32* [[CONV]], align 4
-// CHECK17-NEXT:    [[TMP20:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP22]], align 8
-// CHECK17-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP24]], align 8
-// CHECK17-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP25]], align 8
-// CHECK17-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP20]], i64* [[TMP27]], align 8
-// CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP20]], i64* [[TMP29]], align 8
-// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP30]], align 8
-// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, i32 2, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK17-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK17:       omp_offload.failed11:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i64 [[TMP20]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK17:       omp_offload.cont12:
-// CHECK17-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP36]], align 8
-// CHECK17-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 [[TMP35]], i32* [[CONV]], align 4
+// CHECK17-NEXT:    [[TMP36:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK17-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [10 x i32]**
 // CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP38]], align 8
-// CHECK17-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP39]], align 8
-// CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP40]], align 8
+// CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP41]], align 8
+// CHECK17-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP36]], i64* [[TMP43]], align 8
+// CHECK17-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP36]], i64* [[TMP45]], align 8
+// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP46]], align 8
+// CHECK17-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, i32 1, i8** [[TMP40]], i8** [[TMP41]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK17-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK17:       omp_offload.failed17:
+// CHECK17-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP49]], align 4
+// CHECK17-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 2, i32* [[TMP50]], align 4
+// CHECK17-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 8
+// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP48]], i8*** [[TMP52]], align 8
+// CHECK17-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP54]], align 8
+// CHECK17-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK17-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP56]], align 8
+// CHECK17-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK17-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
+// CHECK17-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK17:       omp_offload.failed13:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i64 [[TMP36]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK17:       omp_offload.cont14:
+// CHECK17-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP60]], align 8
+// CHECK17-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP62]], align 8
+// CHECK17-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP63]], align 8
+// CHECK17-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
+// CHECK17-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK17-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK17-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 8
+// CHECK17-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 8
+// CHECK17-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP70]], align 8
+// CHECK17-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP71]], align 8
+// CHECK17-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP72]], align 8
+// CHECK17-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK17-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK17-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK17-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK17:       omp_offload.failed20:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK17:       omp_offload.cont18:
-// CHECK17-NEXT:    [[TMP44:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    store i32 [[TMP44]], i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK17-NEXT:    [[TMP45:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK17-NEXT:    [[CONV21:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED20]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP45]], i32* [[CONV21]], align 4
-// CHECK17-NEXT:    [[TMP46:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED20]], align 8
-// CHECK17-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP48]], align 8
-// CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP50]], align 8
-// CHECK17-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP46]], i64* [[TMP53]], align 8
-// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP46]], i64* [[TMP55]], align 8
-// CHECK17-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP56]], align 8
-// CHECK17-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK17:       omp_offload.cont21:
+// CHECK17-NEXT:    [[TMP76:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    store i32 [[TMP76]], i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK17-NEXT:    [[TMP77:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK17-NEXT:    [[CONV24:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED23]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP77]], i32* [[CONV24]], align 4
+// CHECK17-NEXT:    [[TMP78:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED23]], align 8
+// CHECK17-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP80]], align 8
+// CHECK17-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP82]], align 8
+// CHECK17-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP83]], align 8
+// CHECK17-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP78]], i64* [[TMP85]], align 8
+// CHECK17-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP78]], i64* [[TMP87]], align 8
+// CHECK17-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS27]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP88]], align 8
+// CHECK17-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS25]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS26]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, i32 2, i8** [[TMP57]], i8** [[TMP58]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK17-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED26:%.*]], label [[OMP_OFFLOAD_CONT27:%.*]]
-// CHECK17:       omp_offload.failed26:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i64 [[TMP46]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT27]]
-// CHECK17:       omp_offload.cont27:
+// CHECK17-NEXT:    [[KERNEL_ARGS29:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP91]], align 4
+// CHECK17-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 2, i32* [[TMP92]], align 4
+// CHECK17-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP89]], i8*** [[TMP93]], align 8
+// CHECK17-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP90]], i8*** [[TMP94]], align 8
+// CHECK17-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP95]], align 8
+// CHECK17-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP96]], align 8
+// CHECK17-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP97]], align 8
+// CHECK17-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP98]], align 8
+// CHECK17-NEXT:    [[TMP99:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]])
+// CHECK17-NEXT:    [[TMP100:%.*]] = icmp ne i32 [[TMP99]], 0
+// CHECK17-NEXT:    br i1 [[TMP100]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
+// CHECK17:       omp_offload.failed30:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i64 [[TMP78]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
+// CHECK17:       omp_offload.cont31:
 // CHECK17-NEXT:    ret i32 0
 //
 //
@@ -12470,34 +13320,34 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[_TMP8:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_17:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[N_CASTED19:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
-// CHECK19-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES23:%.*]] = alloca [4 x i64], align 4
+// CHECK19-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[N_CASTED32:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [3 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [3 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [3 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES36:%.*]] = alloca [3 x i64], align 4
-// CHECK19-NEXT:    [[_TMP37:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_38:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_39:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_46:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[N_CASTED47:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED48:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS49:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS50:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS51:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES52:%.*]] = alloca [4 x i64], align 4
-// CHECK19-NEXT:    [[_TMP53:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_54:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_55:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[N_CASTED34:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS35:%.*]] = alloca [3 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS36:%.*]] = alloca [3 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS37:%.*]] = alloca [3 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES38:%.*]] = alloca [3 x i64], align 4
+// CHECK19-NEXT:    [[_TMP39:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_40:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_41:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_49:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[N_CASTED50:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED51:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS52:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS53:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS54:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES55:%.*]] = alloca [4 x i64], align 4
+// CHECK19-NEXT:    [[_TMP56:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_57:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_58:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK19-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK19-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -12555,262 +13405,347 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK19-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK19-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK19-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK19-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK19-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK19-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK19-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK19-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK19-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
-// CHECK19-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK19-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK19-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK19-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
-// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK19-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK19-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK19-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK19-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK19-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK19-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK19-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK19-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK19-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK19-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK19-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK19-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
+// CHECK19-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK19-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK19-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK19-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK19-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK19-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK19-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK19-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK19-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK19-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK19-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK19-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK19-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK19-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK19-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP59]])
-// CHECK19-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK19-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK19:       omp_offload.failed15:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK19:       omp_offload.cont16:
-// CHECK19-NEXT:    [[TMP62:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP62]], i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK19-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP63]], i32* [[N_CASTED18]], align 4
-// CHECK19-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N_CASTED18]], align 4
-// CHECK19-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK19-NEXT:    store i32 [[TMP65]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK19-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK19-NEXT:    [[TMP67:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP68:%.*]] = sext i32 [[TMP67]] to i64
-// CHECK19-NEXT:    [[TMP69:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP69]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
-// CHECK19-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP64]], i32* [[TMP71]], align 4
-// CHECK19-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP64]], i32* [[TMP73]], align 4
-// CHECK19-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP74]], align 4
-// CHECK19-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP76]], align 4
-// CHECK19-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP78]], align 4
-// CHECK19-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP79]], align 4
-// CHECK19-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 4
-// CHECK19-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 4
-// CHECK19-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 2
-// CHECK19-NEXT:    store i64 [[TMP68]], i64* [[TMP84]], align 4
-// CHECK19-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP85]], align 4
-// CHECK19-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK19-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK19-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP67]])
+// CHECK19-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK19-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK19-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK19-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK19-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK19-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK19-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK19-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK19-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK19-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK19-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK19:       omp_offload.failed16:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l143(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK19:       omp_offload.cont17:
+// CHECK19-NEXT:    [[TMP78:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP78]], i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK19-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP79]], i32* [[N_CASTED19]], align 4
+// CHECK19-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_CASTED19]], align 4
+// CHECK19-NEXT:    [[TMP81:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK19-NEXT:    store i32 [[TMP81]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK19-NEXT:    [[TMP82:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK19-NEXT:    [[TMP83:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP84:%.*]] = sext i32 [[TMP83]] to i64
+// CHECK19-NEXT:    [[TMP85:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES23]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP85]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
+// CHECK19-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP66]], i32* [[TMP87]], align 4
-// CHECK19-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK19-NEXT:    store i32 [[TMP80]], i32* [[TMP87]], align 4
+// CHECK19-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP66]], i32* [[TMP89]], align 4
-// CHECK19-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK19-NEXT:    store i32 [[TMP80]], i32* [[TMP89]], align 4
+// CHECK19-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
 // CHECK19-NEXT:    store i8* null, i8** [[TMP90]], align 4
-// CHECK19-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP94:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP94]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK19-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK19-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP95]], 0
-// CHECK19-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
-// CHECK19-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
-// CHECK19-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK19-NEXT:    [[TMP96:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK19-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP96]], 1
-// CHECK19-NEXT:    [[TMP97:%.*]] = zext i32 [[ADD29]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP97]])
-// CHECK19-NEXT:    [[TMP98:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, i32 4, i8** [[TMP91]], i8** [[TMP92]], i64* [[TMP93]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP99:%.*]] = icmp ne i32 [[TMP98]], 0
-// CHECK19-NEXT:    br i1 [[TMP99]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK19:       omp_offload.failed30:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i32 [[TMP64]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP66]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK19:       omp_offload.cont31:
-// CHECK19-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP100]], i32* [[N_CASTED32]], align 4
-// CHECK19-NEXT:    [[TMP101:%.*]] = load i32, i32* [[N_CASTED32]], align 4
-// CHECK19-NEXT:    [[TMP102:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP103:%.*]] = sext i32 [[TMP102]] to i64
-// CHECK19-NEXT:    [[TMP104:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES36]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP104]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
-// CHECK19-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP106:%.*]] = bitcast i8** [[TMP105]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP101]], i32* [[TMP106]], align 4
-// CHECK19-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP101]], i32* [[TMP108]], align 4
-// CHECK19-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP109]], align 4
-// CHECK19-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP111:%.*]] = bitcast i8** [[TMP110]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP111]], align 4
-// CHECK19-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP113]], align 4
-// CHECK19-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP114]], align 4
-// CHECK19-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP116]], align 4
-// CHECK19-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP118]], align 4
-// CHECK19-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES36]], i32 0, i32 2
-// CHECK19-NEXT:    store i64 [[TMP103]], i64* [[TMP119]], align 4
-// CHECK19-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP120]], align 4
-// CHECK19-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP123:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES36]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP92]], align 4
+// CHECK19-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP94]], align 4
+// CHECK19-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP95]], align 4
+// CHECK19-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 4
+// CHECK19-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 4
+// CHECK19-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 2
+// CHECK19-NEXT:    store i64 [[TMP84]], i64* [[TMP100]], align 4
+// CHECK19-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP101]], align 4
+// CHECK19-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP82]], i32* [[TMP103]], align 4
+// CHECK19-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP82]], i32* [[TMP105]], align 4
+// CHECK19-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 3
+// CHECK19-NEXT:    store i8* null, i8** [[TMP106]], align 4
+// CHECK19-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP110:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP110]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK19-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK19-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP111]], 0
+// CHECK19-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
+// CHECK19-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
+// CHECK19-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK19-NEXT:    [[TMP112:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK19-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP112]], 1
+// CHECK19-NEXT:    [[TMP113:%.*]] = zext i32 [[ADD30]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP113]])
+// CHECK19-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP114]], align 4
+// CHECK19-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 4, i32* [[TMP115]], align 4
+// CHECK19-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK19-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP108]], i8*** [[TMP117]], align 4
+// CHECK19-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP109]], i64** [[TMP118]], align 4
+// CHECK19-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP119]], align 4
+// CHECK19-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK19-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP121]], align 4
+// CHECK19-NEXT:    [[TMP122:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK19-NEXT:    [[TMP123:%.*]] = icmp ne i32 [[TMP122]], 0
+// CHECK19-NEXT:    br i1 [[TMP123]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK19:       omp_offload.failed32:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l147(i32 [[TMP80]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP82]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK19:       omp_offload.cont33:
 // CHECK19-NEXT:    [[TMP124:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP124]], i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK19-NEXT:    [[TMP125:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK19-NEXT:    [[SUB40:%.*]] = sub nsw i32 [[TMP125]], 0
-// CHECK19-NEXT:    [[DIV41:%.*]] = sdiv i32 [[SUB40]], 1
-// CHECK19-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[DIV41]], 1
-// CHECK19-NEXT:    store i32 [[SUB42]], i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK19-NEXT:    [[TMP126:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_39]], align 4
-// CHECK19-NEXT:    [[ADD43:%.*]] = add nsw i32 [[TMP126]], 1
-// CHECK19-NEXT:    [[TMP127:%.*]] = zext i32 [[ADD43]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP127]])
-// CHECK19-NEXT:    [[TMP128:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, i32 3, i8** [[TMP121]], i8** [[TMP122]], i64* [[TMP123]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP129:%.*]] = icmp ne i32 [[TMP128]], 0
-// CHECK19-NEXT:    br i1 [[TMP129]], label [[OMP_OFFLOAD_FAILED44:%.*]], label [[OMP_OFFLOAD_CONT45:%.*]]
-// CHECK19:       omp_offload.failed44:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i32 [[TMP101]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT45]]
-// CHECK19:       omp_offload.cont45:
-// CHECK19-NEXT:    [[TMP130:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP130]], i32* [[DOTCAPTURE_EXPR_46]], align 4
-// CHECK19-NEXT:    [[TMP131:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP131]], i32* [[N_CASTED47]], align 4
-// CHECK19-NEXT:    [[TMP132:%.*]] = load i32, i32* [[N_CASTED47]], align 4
-// CHECK19-NEXT:    [[TMP133:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_46]], align 4
-// CHECK19-NEXT:    store i32 [[TMP133]], i32* [[DOTCAPTURE_EXPR__CASTED48]], align 4
-// CHECK19-NEXT:    [[TMP134:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED48]], align 4
-// CHECK19-NEXT:    [[TMP135:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP136:%.*]] = sext i32 [[TMP135]] to i64
-// CHECK19-NEXT:    [[TMP137:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES52]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP137]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
-// CHECK19-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP132]], i32* [[TMP139]], align 4
-// CHECK19-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP132]], i32* [[TMP141]], align 4
-// CHECK19-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP142]], align 4
-// CHECK19-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP144]], align 4
-// CHECK19-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP146]], align 4
-// CHECK19-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP147]], align 4
-// CHECK19-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP149]], align 4
-// CHECK19-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP151]], align 4
-// CHECK19-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES52]], i32 0, i32 2
-// CHECK19-NEXT:    store i64 [[TMP136]], i64* [[TMP152]], align 4
-// CHECK19-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP153]], align 4
-// CHECK19-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP155:%.*]] = bitcast i8** [[TMP154]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP134]], i32* [[TMP155]], align 4
-// CHECK19-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP134]], i32* [[TMP157]], align 4
-// CHECK19-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS51]], i32 0, i32 3
-// CHECK19-NEXT:    store i8* null, i8** [[TMP158]], align 4
-// CHECK19-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS49]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS50]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES52]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP162:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP162]], i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK19-NEXT:    [[TMP163:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_54]], align 4
-// CHECK19-NEXT:    [[SUB56:%.*]] = sub nsw i32 [[TMP163]], 0
-// CHECK19-NEXT:    [[DIV57:%.*]] = sdiv i32 [[SUB56]], 1
-// CHECK19-NEXT:    [[SUB58:%.*]] = sub nsw i32 [[DIV57]], 1
-// CHECK19-NEXT:    store i32 [[SUB58]], i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK19-NEXT:    [[TMP164:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_55]], align 4
-// CHECK19-NEXT:    [[ADD59:%.*]] = add nsw i32 [[TMP164]], 1
-// CHECK19-NEXT:    [[TMP165:%.*]] = zext i32 [[ADD59]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP165]])
-// CHECK19-NEXT:    [[TMP166:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, i32 4, i8** [[TMP159]], i8** [[TMP160]], i64* [[TMP161]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP167:%.*]] = icmp ne i32 [[TMP166]], 0
-// CHECK19-NEXT:    br i1 [[TMP167]], label [[OMP_OFFLOAD_FAILED60:%.*]], label [[OMP_OFFLOAD_CONT61:%.*]]
-// CHECK19:       omp_offload.failed60:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i32 [[TMP132]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP134]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT61]]
-// CHECK19:       omp_offload.cont61:
-// CHECK19-NEXT:    [[TMP168:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK19-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP168]])
+// CHECK19-NEXT:    store i32 [[TMP124]], i32* [[N_CASTED34]], align 4
+// CHECK19-NEXT:    [[TMP125:%.*]] = load i32, i32* [[N_CASTED34]], align 4
+// CHECK19-NEXT:    [[TMP126:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP127:%.*]] = sext i32 [[TMP126]] to i64
+// CHECK19-NEXT:    [[TMP128:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES38]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP128]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
+// CHECK19-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP125]], i32* [[TMP130]], align 4
+// CHECK19-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP132:%.*]] = bitcast i8** [[TMP131]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP125]], i32* [[TMP132]], align 4
+// CHECK19-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS37]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP133]], align 4
+// CHECK19-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP135]], align 4
+// CHECK19-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP137]], align 4
+// CHECK19-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS37]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP138]], align 4
+// CHECK19-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP140]], align 4
+// CHECK19-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP142]], align 4
+// CHECK19-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES38]], i32 0, i32 2
+// CHECK19-NEXT:    store i64 [[TMP127]], i64* [[TMP143]], align 4
+// CHECK19-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS37]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP144]], align 4
+// CHECK19-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS35]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS36]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES38]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP148:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP148]], i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK19-NEXT:    [[TMP149:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK19-NEXT:    [[SUB42:%.*]] = sub nsw i32 [[TMP149]], 0
+// CHECK19-NEXT:    [[DIV43:%.*]] = sdiv i32 [[SUB42]], 1
+// CHECK19-NEXT:    [[SUB44:%.*]] = sub nsw i32 [[DIV43]], 1
+// CHECK19-NEXT:    store i32 [[SUB44]], i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK19-NEXT:    [[TMP150:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_41]], align 4
+// CHECK19-NEXT:    [[ADD45:%.*]] = add nsw i32 [[TMP150]], 1
+// CHECK19-NEXT:    [[TMP151:%.*]] = zext i32 [[ADD45]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP151]])
+// CHECK19-NEXT:    [[KERNEL_ARGS46:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP152]], align 4
+// CHECK19-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP153]], align 4
+// CHECK19-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP145]], i8*** [[TMP154]], align 4
+// CHECK19-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP146]], i8*** [[TMP155]], align 4
+// CHECK19-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP147]], i64** [[TMP156]], align 4
+// CHECK19-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP157]], align 4
+// CHECK19-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP158]], align 4
+// CHECK19-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP159]], align 4
+// CHECK19-NEXT:    [[TMP160:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS46]])
+// CHECK19-NEXT:    [[TMP161:%.*]] = icmp ne i32 [[TMP160]], 0
+// CHECK19-NEXT:    br i1 [[TMP161]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
+// CHECK19:       omp_offload.failed47:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l151(i32 [[TMP125]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
+// CHECK19:       omp_offload.cont48:
+// CHECK19-NEXT:    [[TMP162:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP162]], i32* [[DOTCAPTURE_EXPR_49]], align 4
+// CHECK19-NEXT:    [[TMP163:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP163]], i32* [[N_CASTED50]], align 4
+// CHECK19-NEXT:    [[TMP164:%.*]] = load i32, i32* [[N_CASTED50]], align 4
+// CHECK19-NEXT:    [[TMP165:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_49]], align 4
+// CHECK19-NEXT:    store i32 [[TMP165]], i32* [[DOTCAPTURE_EXPR__CASTED51]], align 4
+// CHECK19-NEXT:    [[TMP166:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED51]], align 4
+// CHECK19-NEXT:    [[TMP167:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP168:%.*]] = sext i32 [[TMP167]] to i64
+// CHECK19-NEXT:    [[TMP169:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES55]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP169]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
+// CHECK19-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP171:%.*]] = bitcast i8** [[TMP170]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP164]], i32* [[TMP171]], align 4
+// CHECK19-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP173:%.*]] = bitcast i8** [[TMP172]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP164]], i32* [[TMP173]], align 4
+// CHECK19-NEXT:    [[TMP174:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP174]], align 4
+// CHECK19-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP176:%.*]] = bitcast i8** [[TMP175]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP176]], align 4
+// CHECK19-NEXT:    [[TMP177:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP178:%.*]] = bitcast i8** [[TMP177]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP178]], align 4
+// CHECK19-NEXT:    [[TMP179:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP179]], align 4
+// CHECK19-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP181]], align 4
+// CHECK19-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP183:%.*]] = bitcast i8** [[TMP182]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP183]], align 4
+// CHECK19-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES55]], i32 0, i32 2
+// CHECK19-NEXT:    store i64 [[TMP168]], i64* [[TMP184]], align 4
+// CHECK19-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP185]], align 4
+// CHECK19-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP187:%.*]] = bitcast i8** [[TMP186]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP166]], i32* [[TMP187]], align 4
+// CHECK19-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP189:%.*]] = bitcast i8** [[TMP188]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP166]], i32* [[TMP189]], align 4
+// CHECK19-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS54]], i32 0, i32 3
+// CHECK19-NEXT:    store i8* null, i8** [[TMP190]], align 4
+// CHECK19-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS52]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS53]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES55]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP194:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP194]], i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK19-NEXT:    [[TMP195:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_57]], align 4
+// CHECK19-NEXT:    [[SUB59:%.*]] = sub nsw i32 [[TMP195]], 0
+// CHECK19-NEXT:    [[DIV60:%.*]] = sdiv i32 [[SUB59]], 1
+// CHECK19-NEXT:    [[SUB61:%.*]] = sub nsw i32 [[DIV60]], 1
+// CHECK19-NEXT:    store i32 [[SUB61]], i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK19-NEXT:    [[TMP196:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_58]], align 4
+// CHECK19-NEXT:    [[ADD62:%.*]] = add nsw i32 [[TMP196]], 1
+// CHECK19-NEXT:    [[TMP197:%.*]] = zext i32 [[ADD62]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP197]])
+// CHECK19-NEXT:    [[KERNEL_ARGS63:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP198]], align 4
+// CHECK19-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 4, i32* [[TMP199]], align 4
+// CHECK19-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP191]], i8*** [[TMP200]], align 4
+// CHECK19-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP192]], i8*** [[TMP201]], align 4
+// CHECK19-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP193]], i64** [[TMP202]], align 4
+// CHECK19-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP203]], align 4
+// CHECK19-NEXT:    [[TMP204:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP204]], align 4
+// CHECK19-NEXT:    [[TMP205:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP205]], align 4
+// CHECK19-NEXT:    [[TMP206:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS63]])
+// CHECK19-NEXT:    [[TMP207:%.*]] = icmp ne i32 [[TMP206]], 0
+// CHECK19-NEXT:    br i1 [[TMP207]], label [[OMP_OFFLOAD_FAILED64:%.*]], label [[OMP_OFFLOAD_CONT65:%.*]]
+// CHECK19:       omp_offload.failed64:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l155(i32 [[TMP164]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP166]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT65]]
+// CHECK19:       omp_offload.cont65:
+// CHECK19-NEXT:    [[TMP208:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK19-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP208]])
 // CHECK19-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK19-NEXT:    [[TMP169:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP169]])
-// CHECK19-NEXT:    [[TMP170:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK19-NEXT:    ret i32 [[TMP170]]
+// CHECK19-NEXT:    [[TMP209:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP209]])
+// CHECK19-NEXT:    [[TMP210:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK19-NEXT:    ret i32 [[TMP210]]
 //
 //
 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l139
@@ -14055,20 +14990,20 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 4
-// CHECK19-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED20:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 4
+// CHECK19-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR__CASTED23:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK19-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK19-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -14082,111 +15017,196 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK19-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK19-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK19-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK19-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK19-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK19-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK19-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK19-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l112([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
-// CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK19-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK19-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK19-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK19-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK19:       omp_offload.failed5:
+// CHECK19-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK19-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK19-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK19-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK19-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK19:       omp_offload.failed6:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK19:       omp_offload.cont6:
-// CHECK19-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP18]], i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK19-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK19-NEXT:    store i32 [[TMP19]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK19-NEXT:    [[TMP20:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP22]], align 4
-// CHECK19-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP24]], align 4
-// CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP25]], align 4
-// CHECK19-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP20]], i32* [[TMP27]], align 4
-// CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP20]], i32* [[TMP29]], align 4
-// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP30]], align 4
-// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, i32 2, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK19-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK19:       omp_offload.failed11:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i32 [[TMP20]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK19:       omp_offload.cont12:
-// CHECK19-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP36]], align 4
-// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK19:       omp_offload.cont7:
+// CHECK19-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP34]], i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK19-NEXT:    [[TMP35:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
+// CHECK19-NEXT:    store i32 [[TMP35]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK19-NEXT:    [[TMP36:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [10 x i32]**
 // CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP38]], align 4
-// CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP39]], align 4
-// CHECK19-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP40]], align 4
+// CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP41]], align 4
+// CHECK19-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP36]], i32* [[TMP43]], align 4
+// CHECK19-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP36]], i32* [[TMP45]], align 4
+// CHECK19-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP46]], align 4
+// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
+// CHECK19-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP49]], align 4
+// CHECK19-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 2, i32* [[TMP50]], align 4
+// CHECK19-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 4
+// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP48]], i8*** [[TMP52]], align 4
+// CHECK19-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP53]], align 4
+// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP54]], align 4
+// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP55]], align 4
+// CHECK19-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP56]], align 4
+// CHECK19-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK19-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
+// CHECK19-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK19:       omp_offload.failed13:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l120([10 x i32]* [[A]], i32 [[TMP36]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK19:       omp_offload.cont14:
+// CHECK19-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP60:%.*]] = bitcast i8** [[TMP59]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP60]], align 4
+// CHECK19-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP62:%.*]] = bitcast i8** [[TMP61]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP62]], align 4
+// CHECK19-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP63]], align 4
+// CHECK19-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, i32 1, i8** [[TMP40]], i8** [[TMP41]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK19-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK19:       omp_offload.failed17:
+// CHECK19-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK19-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK19-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK19-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 4
+// CHECK19-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK19-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP71]], align 4
+// CHECK19-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK19-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP73]], align 4
+// CHECK19-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK19-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK19-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK19:       omp_offload.failed20:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l124([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK19:       omp_offload.cont18:
-// CHECK19-NEXT:    [[TMP44:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP44]], i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK19-NEXT:    [[TMP45:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK19-NEXT:    store i32 [[TMP45]], i32* [[DOTCAPTURE_EXPR__CASTED20]], align 4
-// CHECK19-NEXT:    [[TMP46:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED20]], align 4
-// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP48]], align 4
-// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP50]], align 4
-// CHECK19-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP51]], align 4
-// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP46]], i32* [[TMP53]], align 4
-// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP46]], i32* [[TMP55]], align 4
-// CHECK19-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP56]], align 4
-// CHECK19-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK19:       omp_offload.cont21:
+// CHECK19-NEXT:    [[TMP76:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP76]], i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK19-NEXT:    [[TMP77:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK19-NEXT:    store i32 [[TMP77]], i32* [[DOTCAPTURE_EXPR__CASTED23]], align 4
+// CHECK19-NEXT:    [[TMP78:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED23]], align 4
+// CHECK19-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP80]], align 4
+// CHECK19-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP82]], align 4
+// CHECK19-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP83]], align 4
+// CHECK19-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP78]], i32* [[TMP85]], align 4
+// CHECK19-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP78]], i32* [[TMP87]], align 4
+// CHECK19-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP88]], align 4
+// CHECK19-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, i32 2, i8** [[TMP57]], i8** [[TMP58]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK19-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
-// CHECK19:       omp_offload.failed25:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i32 [[TMP46]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
-// CHECK19:       omp_offload.cont26:
+// CHECK19-NEXT:    [[KERNEL_ARGS28:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP91]], align 4
+// CHECK19-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 2, i32* [[TMP92]], align 4
+// CHECK19-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP89]], i8*** [[TMP93]], align 4
+// CHECK19-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP90]], i8*** [[TMP94]], align 4
+// CHECK19-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP95]], align 4
+// CHECK19-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP96]], align 4
+// CHECK19-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP97]], align 4
+// CHECK19-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP98]], align 4
+// CHECK19-NEXT:    [[TMP99:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]])
+// CHECK19-NEXT:    [[TMP100:%.*]] = icmp ne i32 [[TMP99]], 0
+// CHECK19-NEXT:    br i1 [[TMP100]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
+// CHECK19:       omp_offload.failed29:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l128([10 x i32]* [[A]], i32 [[TMP78]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
+// CHECK19:       omp_offload.cont30:
 // CHECK19-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_uses_allocators_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_uses_allocators_codegen.cpp
index 7c00f34..2ac4423 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_uses_allocators_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_simd_uses_allocators_codegen.cpp
@@ -65,7 +65,7 @@ void foo() {
   omp_alloctrait_t traits[10];
   omp_allocator_handle_t my_allocator;
 
-// CHECK: [[RES:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @.[[TGT_REGION:.+]].region_id, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZES]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPTYPES]], i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
+// CHECK: [[RES:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 0, i32 0, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
 // CHECK: [[CMP:%.+]] = icmp ne i32 [[RES]], 0
 // CHECK: br i1 [[CMP]], label %[[FAILED:.+]], label %[[DONE:.+]]
 // CHECK: [[FAILED]]:

diff --git a/clang/test/OpenMP/target_teams_distribute_parallel_for_uses_allocators_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_parallel_for_uses_allocators_codegen.cpp
index 7ac0591..b87fecf 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_parallel_for_uses_allocators_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_parallel_for_uses_allocators_codegen.cpp
@@ -65,7 +65,7 @@ void foo() {
   omp_alloctrait_t traits[10];
   omp_allocator_handle_t my_allocator;
 
-// CHECK: [[RES:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @.[[TGT_REGION:.+]].region_id, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZES]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPTYPES]], i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
+// CHECK: [[RES:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 0, i32 0, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
 // CHECK: [[CMP:%.+]] = icmp ne i32 [[RES]], 0
 // CHECK: br i1 [[CMP]], label %[[FAILED:.+]], label %[[DONE:.+]]
 // CHECK: [[FAILED]]:

diff --git a/clang/test/OpenMP/target_teams_distribute_private_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_private_codegen.cpp
index e990f51..210fdcc 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_private_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_private_codegen.cpp
@@ -239,9 +239,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l91.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l91.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l91() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -385,27 +402,44 @@ int main() {
 // CHECK1-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 8
 // CHECK1-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 8
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK1-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP4]]
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev
@@ -727,9 +761,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l91.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l91.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l91() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -871,27 +922,44 @@ int main() {
 // CHECK3-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 4
 // CHECK3-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK3-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP4]]
+// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev

diff --git a/clang/test/OpenMP/target_teams_distribute_reduction_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_reduction_codegen.cpp
index aa9c5c8..c4d8f49 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_reduction_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_reduction_codegen.cpp
@@ -98,9 +98,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l60.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l60.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l60(i32* @_ZZ4mainE5sivar) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -258,9 +275,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
-// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
+// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i32* [[T_VAR]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -421,9 +455,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l60.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l60.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l60(i32* @_ZZ4mainE5sivar) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -581,9 +632,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
-// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
+// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i32* [[T_VAR]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/target_teams_distribute_simd_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_simd_codegen.cpp
index e3cda79..31b04d9 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_simd_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_simd_codegen.cpp
@@ -326,12 +326,12 @@ int bar(int n){
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS17:%.*]] = alloca [2 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS18:%.*]] = alloca [2 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP19:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[A_CASTED22:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS26:%.*]] = alloca [9 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS27:%.*]] = alloca [9 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS28:%.*]] = alloca [9 x i8*], align 8
+// CHECK1-NEXT:    [[A_CASTED23:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [9 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [9 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [9 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 8
-// CHECK1-NEXT:    [[_TMP29:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK1-NEXT:    store i32 0, i32* [[A]], align 4
@@ -442,165 +442,216 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK1-NEXT:    [[TMP68:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP66]], i8** [[TMP67]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP69:%.*]] = icmp ne i32 [[TMP68]], 0
-// CHECK1-NEXT:    br i1 [[TMP69]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK1-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP69]], align 4
+// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP66]], i8*** [[TMP70]], align 8
+// CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP67]], i8*** [[TMP71]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 8
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP75]], align 8
+// CHECK1-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK1-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i64 [[TMP60]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP70:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP78:%.*]] = load i32, i32* [[A]], align 4
 // CHECK1-NEXT:    [[CONV13:%.*]] = bitcast i64* [[A_CASTED12]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP70]], i32* [[CONV13]], align 4
-// CHECK1-NEXT:    [[TMP71:%.*]] = load i64, i64* [[A_CASTED12]], align 8
-// CHECK1-NEXT:    [[TMP72:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK1-NEXT:    store i32 [[TMP78]], i32* [[CONV13]], align 4
+// CHECK1-NEXT:    [[TMP79:%.*]] = load i64, i64* [[A_CASTED12]], align 8
+// CHECK1-NEXT:    [[TMP80:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK1-NEXT:    [[CONV15:%.*]] = bitcast i64* [[AA_CASTED14]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP72]], i16* [[CONV15]], align 2
-// CHECK1-NEXT:    [[TMP73:%.*]] = load i64, i64* [[AA_CASTED14]], align 8
-// CHECK1-NEXT:    [[TMP74:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP74]], 10
+// CHECK1-NEXT:    store i16 [[TMP80]], i16* [[CONV15]], align 2
+// CHECK1-NEXT:    [[TMP81:%.*]] = load i64, i64* [[AA_CASTED14]], align 8
+// CHECK1-NEXT:    [[TMP82:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP82]], 10
 // CHECK1-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK1:       omp_if.then:
-// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP71]], i64* [[TMP76]], align 8
-// CHECK1-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP71]], i64* [[TMP78]], align 8
-// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP79]], align 8
-// CHECK1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP73]], i64* [[TMP81]], align 8
-// CHECK1-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP73]], i64* [[TMP83]], align 8
-// CHECK1-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP84]], align 8
-// CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP79]], i64* [[TMP84]], align 8
+// CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP79]], i64* [[TMP86]], align 8
+// CHECK1-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP87]], align 8
+// CHECK1-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP81]], i64* [[TMP89]], align 8
+// CHECK1-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP81]], i64* [[TMP91]], align 8
+// CHECK1-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP92]], align 8
+// CHECK1-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK1-NEXT:    [[TMP87:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP85]], i8** [[TMP86]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP88:%.*]] = icmp ne i32 [[TMP87]], 0
-// CHECK1-NEXT:    br i1 [[TMP88]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
-// CHECK1:       omp_offload.failed20:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
-// CHECK1:       omp_offload.cont21:
+// CHECK1-NEXT:    [[KERNEL_ARGS20:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP95]], align 4
+// CHECK1-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP96]], align 4
+// CHECK1-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP93]], i8*** [[TMP97]], align 8
+// CHECK1-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP94]], i8*** [[TMP98]], align 8
+// CHECK1-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP99]], align 8
+// CHECK1-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP100]], align 8
+// CHECK1-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP101]], align 8
+// CHECK1-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP102]], align 8
+// CHECK1-NEXT:    [[TMP103:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]])
+// CHECK1-NEXT:    [[TMP104:%.*]] = icmp ne i32 [[TMP103]], 0
+// CHECK1-NEXT:    br i1 [[TMP104]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
+// CHECK1:       omp_offload.failed21:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP79]], i64 [[TMP81]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
+// CHECK1:       omp_offload.cont22:
 // CHECK1-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK1:       omp_if.else:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR4]]
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP79]], i64 [[TMP81]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP89:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    [[CONV23:%.*]] = bitcast i64* [[A_CASTED22]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP89]], i32* [[CONV23]], align 4
-// CHECK1-NEXT:    [[TMP90:%.*]] = load i64, i64* [[A_CASTED22]], align 8
-// CHECK1-NEXT:    [[TMP91:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[CMP24:%.*]] = icmp sgt i32 [[TMP91]], 20
-// CHECK1-NEXT:    br i1 [[CMP24]], label [[OMP_IF_THEN25:%.*]], label [[OMP_IF_ELSE32:%.*]]
-// CHECK1:       omp_if.then25:
-// CHECK1-NEXT:    [[TMP92:%.*]] = mul nuw i64 [[TMP2]], 4
-// CHECK1-NEXT:    [[TMP93:%.*]] = mul nuw i64 5, [[TMP5]]
-// CHECK1-NEXT:    [[TMP94:%.*]] = mul nuw i64 [[TMP93]], 8
-// CHECK1-NEXT:    [[TMP95:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP95]], i8* align 8 bitcast ([9 x i64]* @.offload_sizes.10 to i8*), i64 72, i1 false)
-// CHECK1-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP90]], i64* [[TMP97]], align 8
-// CHECK1-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP90]], i64* [[TMP99]], align 8
-// CHECK1-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP100]], align 8
-// CHECK1-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
-// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 8
-// CHECK1-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
-// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 8
-// CHECK1-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP105]], align 8
-// CHECK1-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP107]], align 8
-// CHECK1-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP109]], align 8
-// CHECK1-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP110]], align 8
-// CHECK1-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to float**
-// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP112]], align 8
-// CHECK1-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
-// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP114]], align 8
-// CHECK1-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK1-NEXT:    store i64 [[TMP92]], i64* [[TMP115]], align 8
-// CHECK1-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 3
+// CHECK1-NEXT:    [[TMP105:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[CONV24:%.*]] = bitcast i64* [[A_CASTED23]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP105]], i32* [[CONV24]], align 4
+// CHECK1-NEXT:    [[TMP106:%.*]] = load i64, i64* [[A_CASTED23]], align 8
+// CHECK1-NEXT:    [[TMP107:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[CMP25:%.*]] = icmp sgt i32 [[TMP107]], 20
+// CHECK1-NEXT:    br i1 [[CMP25]], label [[OMP_IF_THEN26:%.*]], label [[OMP_IF_ELSE34:%.*]]
+// CHECK1:       omp_if.then26:
+// CHECK1-NEXT:    [[TMP108:%.*]] = mul nuw i64 [[TMP2]], 4
+// CHECK1-NEXT:    [[TMP109:%.*]] = mul nuw i64 5, [[TMP5]]
+// CHECK1-NEXT:    [[TMP110:%.*]] = mul nuw i64 [[TMP109]], 8
+// CHECK1-NEXT:    [[TMP111:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK1-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP111]], i8* align 8 bitcast ([9 x i64]* @.offload_sizes.10 to i8*), i64 72, i1 false)
+// CHECK1-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP106]], i64* [[TMP113]], align 8
+// CHECK1-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP106]], i64* [[TMP115]], align 8
+// CHECK1-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 0
 // CHECK1-NEXT:    store i8* null, i8** [[TMP116]], align 8
-// CHECK1-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to [5 x [10 x double]]**
-// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP118]], align 8
-// CHECK1-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
-// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 8
-// CHECK1-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 4
+// CHECK1-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to [10 x float]**
+// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP118]], align 8
+// CHECK1-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [10 x float]**
+// CHECK1-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP120]], align 8
+// CHECK1-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 1
 // CHECK1-NEXT:    store i8* null, i8** [[TMP121]], align 8
-// CHECK1-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP123:%.*]] = bitcast i8** [[TMP122]] to i64*
-// CHECK1-NEXT:    store i64 5, i64* [[TMP123]], align 8
-// CHECK1-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 5
+// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP123]], align 8
+// CHECK1-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP125:%.*]] = bitcast i8** [[TMP124]] to i64*
-// CHECK1-NEXT:    store i64 5, i64* [[TMP125]], align 8
-// CHECK1-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 5
+// CHECK1-NEXT:    store i64 [[TMP2]], i64* [[TMP125]], align 8
+// CHECK1-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 2
 // CHECK1-NEXT:    store i8* null, i8** [[TMP126]], align 8
-// CHECK1-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP128]], align 8
-// CHECK1-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 6
-// CHECK1-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP130]], align 8
-// CHECK1-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 6
-// CHECK1-NEXT:    store i8* null, i8** [[TMP131]], align 8
-// CHECK1-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to double**
-// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP133]], align 8
-// CHECK1-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 7
-// CHECK1-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to double**
-// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP135]], align 8
-// CHECK1-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK1-NEXT:    store i64 [[TMP94]], i64* [[TMP136]], align 8
-// CHECK1-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 7
+// CHECK1-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to float**
+// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP128]], align 8
+// CHECK1-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to float**
+// CHECK1-NEXT:    store float* [[VLA]], float** [[TMP130]], align 8
+// CHECK1-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK1-NEXT:    store i64 [[TMP108]], i64* [[TMP131]], align 8
+// CHECK1-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 3
+// CHECK1-NEXT:    store i8* null, i8** [[TMP132]], align 8
+// CHECK1-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to [5 x [10 x double]]**
+// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP134]], align 8
+// CHECK1-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to [5 x [10 x double]]**
+// CHECK1-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP136]], align 8
+// CHECK1-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 4
 // CHECK1-NEXT:    store i8* null, i8** [[TMP137]], align 8
-// CHECK1-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to %struct.TT**
-// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP139]], align 8
-// CHECK1-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 8
-// CHECK1-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to %struct.TT**
-// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP141]], align 8
-// CHECK1-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 8
+// CHECK1-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i64*
+// CHECK1-NEXT:    store i64 5, i64* [[TMP139]], align 8
+// CHECK1-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 5
+// CHECK1-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i64*
+// CHECK1-NEXT:    store i64 5, i64* [[TMP141]], align 8
+// CHECK1-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 5
 // CHECK1-NEXT:    store i8* null, i8** [[TMP142]], align 8
-// CHECK1-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 6
+// CHECK1-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP144]], align 8
+// CHECK1-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 6
+// CHECK1-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP146]], align 8
+// CHECK1-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 6
+// CHECK1-NEXT:    store i8* null, i8** [[TMP147]], align 8
+// CHECK1-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to double**
+// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP149]], align 8
+// CHECK1-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 7
+// CHECK1-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to double**
+// CHECK1-NEXT:    store double* [[VLA1]], double** [[TMP151]], align 8
+// CHECK1-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK1-NEXT:    store i64 [[TMP110]], i64* [[TMP152]], align 8
+// CHECK1-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 7
+// CHECK1-NEXT:    store i8* null, i8** [[TMP153]], align 8
+// CHECK1-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP155:%.*]] = bitcast i8** [[TMP154]] to %struct.TT**
+// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP155]], align 8
+// CHECK1-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 8
+// CHECK1-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to %struct.TT**
+// CHECK1-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP157]], align 8
+// CHECK1-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 8
+// CHECK1-NEXT:    store i8* null, i8** [[TMP158]], align 8
+// CHECK1-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK1-NEXT:    [[TMP146:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP143]], i8** [[TMP144]], i64* [[TMP145]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.11, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP147:%.*]] = icmp ne i32 [[TMP146]], 0
-// CHECK1-NEXT:    br i1 [[TMP147]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK1:       omp_offload.failed30:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK1:       omp_offload.cont31:
-// CHECK1-NEXT:    br label [[OMP_IF_END33:%.*]]
-// CHECK1:       omp_if.else32:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
-// CHECK1-NEXT:    br label [[OMP_IF_END33]]
-// CHECK1:       omp_if.end33:
-// CHECK1-NEXT:    [[TMP148:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    [[TMP149:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP149]])
-// CHECK1-NEXT:    ret i32 [[TMP148]]
+// CHECK1-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP162]], align 4
+// CHECK1-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 9, i32* [[TMP163]], align 4
+// CHECK1-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP159]], i8*** [[TMP164]], align 8
+// CHECK1-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP160]], i8*** [[TMP165]], align 8
+// CHECK1-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP161]], i64** [[TMP166]], align 8
+// CHECK1-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.11, i32 0, i32 0), i64** [[TMP167]], align 8
+// CHECK1-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP168]], align 8
+// CHECK1-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP169]], align 8
+// CHECK1-NEXT:    [[TMP170:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK1-NEXT:    [[TMP171:%.*]] = icmp ne i32 [[TMP170]], 0
+// CHECK1-NEXT:    br i1 [[TMP171]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK1:       omp_offload.failed32:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP106]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK1:       omp_offload.cont33:
+// CHECK1-NEXT:    br label [[OMP_IF_END35:%.*]]
+// CHECK1:       omp_if.else34:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP106]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
+// CHECK1-NEXT:    br label [[OMP_IF_END35]]
+// CHECK1:       omp_if.end35:
+// CHECK1-NEXT:    [[TMP172:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP173:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP173]])
+// CHECK1-NEXT:    ret i32 [[TMP172]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97
@@ -743,6 +794,7 @@ int bar(int n){
 // CHECK1-NEXT:    [[AA_CASTED_I:%.*]] = alloca i64, align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i64, align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR__CASTED5_I:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK1-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
 // CHECK1-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -783,23 +835,39 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10) #[[ATTR4]]
-// CHECK1-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 1, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK1-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK1-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP26]], align 4, !noalias !26
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP27]], align 4, !noalias !26
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP28]], align 8, !noalias !26
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP29]], align 8, !noalias !26
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP22]], i64** [[TMP30]], align 8, !noalias !26
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP31]], align 8, !noalias !26
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP32]], align 8, !noalias !26
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP33]], align 8, !noalias !26
+// CHECK1-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP25]], i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK1-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK1-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK1:       omp_offload.failed.i:
-// CHECK1-NEXT:    [[TMP28:%.*]] = load i16, i16* [[TMP16]], align 2
+// CHECK1-NEXT:    [[TMP36:%.*]] = load i16, i16* [[TMP16]], align 2
 // CHECK1-NEXT:    [[CONV_I:%.*]] = bitcast i64* [[AA_CASTED_I]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP28]], i16* [[CONV_I]], align 2, !noalias !26
-// CHECK1-NEXT:    [[TMP29:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !26
-// CHECK1-NEXT:    [[TMP30:%.*]] = load i32, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    store i16 [[TMP36]], i16* [[CONV_I]], align 2, !noalias !26
+// CHECK1-NEXT:    [[TMP37:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !26
+// CHECK1-NEXT:    [[TMP38:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK1-NEXT:    [[CONV4_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED_I]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP30]], i32* [[CONV4_I]], align 4, !noalias !26
-// CHECK1-NEXT:    [[TMP31:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED_I]], align 8, !noalias !26
-// CHECK1-NEXT:    [[TMP32:%.*]] = load i32, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    store i32 [[TMP38]], i32* [[CONV4_I]], align 4, !noalias !26
+// CHECK1-NEXT:    [[TMP39:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED_I]], align 8, !noalias !26
+// CHECK1-NEXT:    [[TMP40:%.*]] = load i32, i32* [[TMP24]], align 4
 // CHECK1-NEXT:    [[CONV6_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED5_I]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP32]], i32* [[CONV6_I]], align 4, !noalias !26
-// CHECK1-NEXT:    [[TMP33:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5_I]], align 8, !noalias !26
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97(i64 [[TMP29]], i64 [[TMP31]], i64 [[TMP33]]) #[[ATTR4]]
+// CHECK1-NEXT:    store i32 [[TMP40]], i32* [[CONV6_I]], align 4, !noalias !26
+// CHECK1-NEXT:    [[TMP41:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5_I]], align 8, !noalias !26
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97(i64 [[TMP37]], i64 [[TMP39]], i64 [[TMP41]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
 // CHECK1:       .omp_outlined..1.exit:
 // CHECK1-NEXT:    ret i32 0
@@ -1376,9 +1444,26 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK1-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK1-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.14, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK1-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1388,15 +1473,15 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215(%struct.S1* [[THIS1]], i64 [[TMP6]], i64 2, i64 [[TMP2]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP42:%.*]] = mul nsw i64 1, [[TMP2]]
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP42]]
+// CHECK1-NEXT:    [[TMP50:%.*]] = mul nsw i64 1, [[TMP2]]
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP50]]
 // CHECK1-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
-// CHECK1-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK1-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP44]]
-// CHECK1-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK1-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK1-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK1-NEXT:    [[ADD4:%.*]] = add nsw i32 [[CONV3]], [[TMP52]]
+// CHECK1-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK1-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK1-NEXT:    ret i32 [[ADD4]]
 //
 //
@@ -1501,9 +1586,26 @@ int bar(int n){
 // CHECK1-NEXT:    [[ADD8:%.*]] = add i32 [[TMP40]], 1
 // CHECK1-NEXT:    [[TMP41:%.*]] = zext i32 [[ADD8]] to i64
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP41]])
-// CHECK1-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK1-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP43]], align 4
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP34]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP35]], i8*** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP49]], align 8
+// CHECK1-NEXT:    [[TMP50:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP51:%.*]] = icmp ne i32 [[TMP50]], 0
+// CHECK1-NEXT:    br i1 [[TMP51]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], i64 [[TMP7]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1513,8 +1615,8 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], i64 [[TMP7]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP44:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP44]]
+// CHECK1-NEXT:    [[TMP52:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP52]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -1572,9 +1674,26 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK1-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.19, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.20, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK1-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.19, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.20, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK1-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1584,8 +1703,8 @@ int bar(int n){
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK1-NEXT:    br label [[OMP_IF_END]]
 // CHECK1:       omp_if.end:
-// CHECK1-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP24]]
+// CHECK1-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215
@@ -2047,12 +2166,12 @@ int bar(int n){
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS13:%.*]] = alloca [2 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS14:%.*]] = alloca [2 x i8*], align 4
 // CHECK3-NEXT:    [[_TMP15:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[A_CASTED18:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [9 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [9 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [9 x i8*], align 4
+// CHECK3-NEXT:    [[A_CASTED19:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [9 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [9 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [9 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 4
-// CHECK3-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK3-NEXT:    store i32 0, i32* [[A]], align 4
@@ -2158,165 +2277,216 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK3-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP64]], i8** [[TMP65]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
-// CHECK3-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK3-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK3-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP71]], align 4
+// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP73]], align 4
+// CHECK3-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK3-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i32 [[TMP58]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP68:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP68]], i32* [[A_CASTED9]], align 4
-// CHECK3-NEXT:    [[TMP69:%.*]] = load i32, i32* [[A_CASTED9]], align 4
-// CHECK3-NEXT:    [[TMP70:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK3-NEXT:    [[TMP76:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP76]], i32* [[A_CASTED9]], align 4
+// CHECK3-NEXT:    [[TMP77:%.*]] = load i32, i32* [[A_CASTED9]], align 4
+// CHECK3-NEXT:    [[TMP78:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK3-NEXT:    [[CONV11:%.*]] = bitcast i32* [[AA_CASTED10]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP70]], i16* [[CONV11]], align 2
-// CHECK3-NEXT:    [[TMP71:%.*]] = load i32, i32* [[AA_CASTED10]], align 4
-// CHECK3-NEXT:    [[TMP72:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP72]], 10
+// CHECK3-NEXT:    store i16 [[TMP78]], i16* [[CONV11]], align 2
+// CHECK3-NEXT:    [[TMP79:%.*]] = load i32, i32* [[AA_CASTED10]], align 4
+// CHECK3-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP80]], 10
 // CHECK3-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK3:       omp_if.then:
-// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP69]], i32* [[TMP74]], align 4
-// CHECK3-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP69]], i32* [[TMP76]], align 4
-// CHECK3-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP77]], align 4
-// CHECK3-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP71]], i32* [[TMP79]], align 4
-// CHECK3-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP71]], i32* [[TMP81]], align 4
-// CHECK3-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP82]], align 4
-// CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP77]], i32* [[TMP82]], align 4
+// CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP77]], i32* [[TMP84]], align 4
+// CHECK3-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP85]], align 4
+// CHECK3-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP79]], i32* [[TMP87]], align 4
+// CHECK3-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP79]], i32* [[TMP89]], align 4
+// CHECK3-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP90]], align 4
+// CHECK3-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK3-NEXT:    [[TMP85:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP83]], i8** [[TMP84]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP86:%.*]] = icmp ne i32 [[TMP85]], 0
-// CHECK3-NEXT:    br i1 [[TMP86]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK3:       omp_offload.failed16:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR4]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP93]], align 4
+// CHECK3-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP94]], align 4
+// CHECK3-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP91]], i8*** [[TMP95]], align 4
+// CHECK3-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP92]], i8*** [[TMP96]], align 4
+// CHECK3-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP97]], align 4
+// CHECK3-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP98]], align 4
+// CHECK3-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP99]], align 4
+// CHECK3-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP100]], align 4
+// CHECK3-NEXT:    [[TMP101:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK3-NEXT:    [[TMP102:%.*]] = icmp ne i32 [[TMP101]], 0
+// CHECK3-NEXT:    br i1 [[TMP102]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK3:       omp_offload.failed17:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP77]], i32 [[TMP79]]) #[[ATTR4]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK3:       omp_offload.cont18:
 // CHECK3-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK3:       omp_if.else:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR4]]
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP77]], i32 [[TMP79]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP87:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP87]], i32* [[A_CASTED18]], align 4
-// CHECK3-NEXT:    [[TMP88:%.*]] = load i32, i32* [[A_CASTED18]], align 4
-// CHECK3-NEXT:    [[TMP89:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[CMP19:%.*]] = icmp sgt i32 [[TMP89]], 20
-// CHECK3-NEXT:    br i1 [[CMP19]], label [[OMP_IF_THEN20:%.*]], label [[OMP_IF_ELSE27:%.*]]
-// CHECK3:       omp_if.then20:
-// CHECK3-NEXT:    [[TMP90:%.*]] = mul nuw i32 [[TMP1]], 4
-// CHECK3-NEXT:    [[TMP91:%.*]] = sext i32 [[TMP90]] to i64
-// CHECK3-NEXT:    [[TMP92:%.*]] = mul nuw i32 5, [[TMP3]]
-// CHECK3-NEXT:    [[TMP93:%.*]] = mul nuw i32 [[TMP92]], 8
-// CHECK3-NEXT:    [[TMP94:%.*]] = sext i32 [[TMP93]] to i64
-// CHECK3-NEXT:    [[TMP95:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP95]], i8* align 4 bitcast ([9 x i64]* @.offload_sizes.10 to i8*), i32 72, i1 false)
-// CHECK3-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP88]], i32* [[TMP97]], align 4
-// CHECK3-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP88]], i32* [[TMP99]], align 4
-// CHECK3-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP100]], align 4
-// CHECK3-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
-// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 4
-// CHECK3-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
-// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 4
-// CHECK3-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP105]], align 4
-// CHECK3-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP107]], align 4
-// CHECK3-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP109]], align 4
-// CHECK3-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP110]], align 4
-// CHECK3-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to float**
-// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP112]], align 4
-// CHECK3-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
-// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP114]], align 4
-// CHECK3-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK3-NEXT:    store i64 [[TMP91]], i64* [[TMP115]], align 4
-// CHECK3-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP103:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP103]], i32* [[A_CASTED19]], align 4
+// CHECK3-NEXT:    [[TMP104:%.*]] = load i32, i32* [[A_CASTED19]], align 4
+// CHECK3-NEXT:    [[TMP105:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[CMP20:%.*]] = icmp sgt i32 [[TMP105]], 20
+// CHECK3-NEXT:    br i1 [[CMP20]], label [[OMP_IF_THEN21:%.*]], label [[OMP_IF_ELSE29:%.*]]
+// CHECK3:       omp_if.then21:
+// CHECK3-NEXT:    [[TMP106:%.*]] = mul nuw i32 [[TMP1]], 4
+// CHECK3-NEXT:    [[TMP107:%.*]] = sext i32 [[TMP106]] to i64
+// CHECK3-NEXT:    [[TMP108:%.*]] = mul nuw i32 5, [[TMP3]]
+// CHECK3-NEXT:    [[TMP109:%.*]] = mul nuw i32 [[TMP108]], 8
+// CHECK3-NEXT:    [[TMP110:%.*]] = sext i32 [[TMP109]] to i64
+// CHECK3-NEXT:    [[TMP111:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK3-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP111]], i8* align 4 bitcast ([9 x i64]* @.offload_sizes.10 to i8*), i32 72, i1 false)
+// CHECK3-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP104]], i32* [[TMP113]], align 4
+// CHECK3-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP104]], i32* [[TMP115]], align 4
+// CHECK3-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 0
 // CHECK3-NEXT:    store i8* null, i8** [[TMP116]], align 4
-// CHECK3-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to [5 x [10 x double]]**
-// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP118]], align 4
-// CHECK3-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
-// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 4
-// CHECK3-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to [10 x float]**
+// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP118]], align 4
+// CHECK3-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [10 x float]**
+// CHECK3-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP120]], align 4
+// CHECK3-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 1
 // CHECK3-NEXT:    store i8* null, i8** [[TMP121]], align 4
-// CHECK3-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP123:%.*]] = bitcast i8** [[TMP122]] to i32*
-// CHECK3-NEXT:    store i32 5, i32* [[TMP123]], align 4
-// CHECK3-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 5
+// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP123]], align 4
+// CHECK3-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP125:%.*]] = bitcast i8** [[TMP124]] to i32*
-// CHECK3-NEXT:    store i32 5, i32* [[TMP125]], align 4
-// CHECK3-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 5
+// CHECK3-NEXT:    store i32 [[TMP1]], i32* [[TMP125]], align 4
+// CHECK3-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 2
 // CHECK3-NEXT:    store i8* null, i8** [[TMP126]], align 4
-// CHECK3-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 6
-// CHECK3-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP128]], align 4
-// CHECK3-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 6
-// CHECK3-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP130]], align 4
-// CHECK3-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 6
-// CHECK3-NEXT:    store i8* null, i8** [[TMP131]], align 4
-// CHECK3-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 7
-// CHECK3-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to double**
-// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP133]], align 4
-// CHECK3-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 7
-// CHECK3-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to double**
-// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP135]], align 4
-// CHECK3-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK3-NEXT:    store i64 [[TMP94]], i64* [[TMP136]], align 4
-// CHECK3-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to float**
+// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP128]], align 4
+// CHECK3-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to float**
+// CHECK3-NEXT:    store float* [[VLA]], float** [[TMP130]], align 4
+// CHECK3-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK3-NEXT:    store i64 [[TMP107]], i64* [[TMP131]], align 4
+// CHECK3-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 3
+// CHECK3-NEXT:    store i8* null, i8** [[TMP132]], align 4
+// CHECK3-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to [5 x [10 x double]]**
+// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP134]], align 4
+// CHECK3-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to [5 x [10 x double]]**
+// CHECK3-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP136]], align 4
+// CHECK3-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 4
 // CHECK3-NEXT:    store i8* null, i8** [[TMP137]], align 4
-// CHECK3-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 8
-// CHECK3-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to %struct.TT**
-// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP139]], align 4
-// CHECK3-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 8
-// CHECK3-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to %struct.TT**
-// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP141]], align 4
-// CHECK3-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32*
+// CHECK3-NEXT:    store i32 5, i32* [[TMP139]], align 4
+// CHECK3-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 5
+// CHECK3-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32*
+// CHECK3-NEXT:    store i32 5, i32* [[TMP141]], align 4
+// CHECK3-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 5
 // CHECK3-NEXT:    store i8* null, i8** [[TMP142]], align 4
-// CHECK3-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 6
+// CHECK3-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP144]], align 4
+// CHECK3-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 6
+// CHECK3-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[TMP146]], align 4
+// CHECK3-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 6
+// CHECK3-NEXT:    store i8* null, i8** [[TMP147]], align 4
+// CHECK3-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to double**
+// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP149]], align 4
+// CHECK3-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 7
+// CHECK3-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to double**
+// CHECK3-NEXT:    store double* [[VLA1]], double** [[TMP151]], align 4
+// CHECK3-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK3-NEXT:    store i64 [[TMP110]], i64* [[TMP152]], align 4
+// CHECK3-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 7
+// CHECK3-NEXT:    store i8* null, i8** [[TMP153]], align 4
+// CHECK3-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP155:%.*]] = bitcast i8** [[TMP154]] to %struct.TT**
+// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP155]], align 4
+// CHECK3-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 8
+// CHECK3-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to %struct.TT**
+// CHECK3-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP157]], align 4
+// CHECK3-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 8
+// CHECK3-NEXT:    store i8* null, i8** [[TMP158]], align 4
+// CHECK3-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK3-NEXT:    [[TMP146:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP143]], i8** [[TMP144]], i64* [[TMP145]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.11, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP147:%.*]] = icmp ne i32 [[TMP146]], 0
-// CHECK3-NEXT:    br i1 [[TMP147]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
-// CHECK3:       omp_offload.failed25:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
-// CHECK3:       omp_offload.cont26:
-// CHECK3-NEXT:    br label [[OMP_IF_END28:%.*]]
-// CHECK3:       omp_if.else27:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
-// CHECK3-NEXT:    br label [[OMP_IF_END28]]
-// CHECK3:       omp_if.end28:
-// CHECK3-NEXT:    [[TMP148:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    [[TMP149:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP149]])
-// CHECK3-NEXT:    ret i32 [[TMP148]]
+// CHECK3-NEXT:    [[KERNEL_ARGS26:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP162]], align 4
+// CHECK3-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 9, i32* [[TMP163]], align 4
+// CHECK3-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP159]], i8*** [[TMP164]], align 4
+// CHECK3-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP160]], i8*** [[TMP165]], align 4
+// CHECK3-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP161]], i64** [[TMP166]], align 4
+// CHECK3-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.11, i32 0, i32 0), i64** [[TMP167]], align 4
+// CHECK3-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP168]], align 4
+// CHECK3-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP169]], align 4
+// CHECK3-NEXT:    [[TMP170:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]])
+// CHECK3-NEXT:    [[TMP171:%.*]] = icmp ne i32 [[TMP170]], 0
+// CHECK3-NEXT:    br i1 [[TMP171]], label [[OMP_OFFLOAD_FAILED27:%.*]], label [[OMP_OFFLOAD_CONT28:%.*]]
+// CHECK3:       omp_offload.failed27:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP104]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT28]]
+// CHECK3:       omp_offload.cont28:
+// CHECK3-NEXT:    br label [[OMP_IF_END30:%.*]]
+// CHECK3:       omp_if.else29:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP104]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
+// CHECK3-NEXT:    br label [[OMP_IF_END30]]
+// CHECK3:       omp_if.end30:
+// CHECK3-NEXT:    [[TMP172:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    [[TMP173:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP173]])
+// CHECK3-NEXT:    ret i32 [[TMP172]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97
@@ -2457,6 +2627,7 @@ int bar(int n){
 // CHECK3-NEXT:    [[AA_CASTED_I:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR__CASTED4_I:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK3-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 4
 // CHECK3-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -2497,21 +2668,37 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
 // CHECK3-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10) #[[ATTR4]]
-// CHECK3-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 1, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK3-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK3-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP26]], align 4, !noalias !27
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP27]], align 4, !noalias !27
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP20]], i8*** [[TMP28]], align 4, !noalias !27
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP29]], align 4, !noalias !27
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP22]], i64** [[TMP30]], align 4, !noalias !27
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP31]], align 4, !noalias !27
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP32]], align 4, !noalias !27
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP33]], align 4, !noalias !27
+// CHECK3-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP25]], i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK3-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK3-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK3:       omp_offload.failed.i:
-// CHECK3-NEXT:    [[TMP28:%.*]] = load i16, i16* [[TMP16]], align 2
+// CHECK3-NEXT:    [[TMP36:%.*]] = load i16, i16* [[TMP16]], align 2
 // CHECK3-NEXT:    [[CONV_I:%.*]] = bitcast i32* [[AA_CASTED_I]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP28]], i16* [[CONV_I]], align 2, !noalias !27
-// CHECK3-NEXT:    [[TMP29:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !27
-// CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[TMP23]], align 4
-// CHECK3-NEXT:    store i32 [[TMP30]], i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !27
-// CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !27
-// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[TMP24]], align 4
-// CHECK3-NEXT:    store i32 [[TMP32]], i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !27
-// CHECK3-NEXT:    [[TMP33:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !27
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97(i32 [[TMP29]], i32 [[TMP31]], i32 [[TMP33]]) #[[ATTR4]]
+// CHECK3-NEXT:    store i16 [[TMP36]], i16* [[CONV_I]], align 2, !noalias !27
+// CHECK3-NEXT:    [[TMP37:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !27
+// CHECK3-NEXT:    [[TMP38:%.*]] = load i32, i32* [[TMP23]], align 4
+// CHECK3-NEXT:    store i32 [[TMP38]], i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !27
+// CHECK3-NEXT:    [[TMP39:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !27
+// CHECK3-NEXT:    [[TMP40:%.*]] = load i32, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    store i32 [[TMP40]], i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !27
+// CHECK3-NEXT:    [[TMP41:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !27
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97(i32 [[TMP37]], i32 [[TMP39]], i32 [[TMP41]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
 // CHECK3:       .omp_outlined..1.exit:
 // CHECK3-NEXT:    ret i32 0
@@ -3078,9 +3265,26 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [5 x i64], [5 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK3-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215.region_id, i32 5, i8** [[TMP37]], i8** [[TMP38]], i64* [[TMP39]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
-// CHECK3-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* [[TMP39]], i64** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.14, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK3-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3090,15 +3294,15 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215(%struct.S1* [[THIS1]], i32 [[TMP5]], i32 2, i32 [[TMP1]], i16* [[VLA]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP42:%.*]] = mul nsw i32 1, [[TMP1]]
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP42]]
+// CHECK3-NEXT:    [[TMP50:%.*]] = mul nsw i32 1, [[TMP1]]
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP50]]
 // CHECK3-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
-// CHECK3-NEXT:    [[TMP43:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
-// CHECK3-NEXT:    [[CONV:%.*]] = sext i16 [[TMP43]] to i32
-// CHECK3-NEXT:    [[TMP44:%.*]] = load i32, i32* [[B]], align 4
-// CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP44]]
-// CHECK3-NEXT:    [[TMP45:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP45]])
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX2]], align 2
+// CHECK3-NEXT:    [[CONV:%.*]] = sext i16 [[TMP51]] to i32
+// CHECK3-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
+// CHECK3-NEXT:    [[ADD3:%.*]] = add nsw i32 [[CONV]], [[TMP52]]
+// CHECK3-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK3-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
 // CHECK3-NEXT:    ret i32 [[ADD3]]
 //
 //
@@ -3201,9 +3405,26 @@ int bar(int n){
 // CHECK3-NEXT:    [[ADD6:%.*]] = add i32 [[TMP40]], 1
 // CHECK3-NEXT:    [[TMP41:%.*]] = zext i32 [[ADD6]] to i64
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP41]])
-// CHECK3-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK3-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP34]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP35]], i8*** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP49]], align 4
+// CHECK3-NEXT:    [[TMP50:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP51:%.*]] = icmp ne i32 [[TMP50]], 0
+// CHECK3-NEXT:    br i1 [[TMP51]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], i32 [[TMP7]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3213,8 +3434,8 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], i32 [[TMP7]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP44:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP44]]
+// CHECK3-NEXT:    [[TMP52:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP52]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -3271,9 +3492,26 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK3-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.19, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.20, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK3-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.19, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.20, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK3-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3283,8 +3521,8 @@ int bar(int n){
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK3-NEXT:    br label [[OMP_IF_END]]
 // CHECK3:       omp_if.end:
-// CHECK3-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP24]]
+// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215
@@ -3734,12 +3972,12 @@ int bar(int n){
 // CHECK5-NEXT:    [[DOTOFFLOAD_PTRS17:%.*]] = alloca [2 x i8*], align 8
 // CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS18:%.*]] = alloca [2 x i8*], align 8
 // CHECK5-NEXT:    [[_TMP19:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[A_CASTED22:%.*]] = alloca i64, align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS26:%.*]] = alloca [9 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS27:%.*]] = alloca [9 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS28:%.*]] = alloca [9 x i8*], align 8
+// CHECK5-NEXT:    [[A_CASTED23:%.*]] = alloca i64, align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [9 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [9 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [9 x i8*], align 8
 // CHECK5-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 8
-// CHECK5-NEXT:    [[_TMP29:%.*]] = alloca i32, align 4
+// CHECK5-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK5-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK5-NEXT:    store i32 0, i32* [[A]], align 4
@@ -3850,165 +4088,216 @@ int bar(int n){
 // CHECK5-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK5-NEXT:    [[TMP68:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP66]], i8** [[TMP67]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP69:%.*]] = icmp ne i32 [[TMP68]], 0
-// CHECK5-NEXT:    br i1 [[TMP69]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK5-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP69]], align 4
+// CHECK5-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP66]], i8*** [[TMP70]], align 8
+// CHECK5-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP67]], i8*** [[TMP71]], align 8
+// CHECK5-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK5-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 8
+// CHECK5-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK5-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP75]], align 8
+// CHECK5-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK5-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i64 [[TMP60]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
-// CHECK5-NEXT:    [[TMP70:%.*]] = load i32, i32* [[A]], align 4
+// CHECK5-NEXT:    [[TMP78:%.*]] = load i32, i32* [[A]], align 4
 // CHECK5-NEXT:    [[CONV13:%.*]] = bitcast i64* [[A_CASTED12]] to i32*
-// CHECK5-NEXT:    store i32 [[TMP70]], i32* [[CONV13]], align 4
-// CHECK5-NEXT:    [[TMP71:%.*]] = load i64, i64* [[A_CASTED12]], align 8
-// CHECK5-NEXT:    [[TMP72:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK5-NEXT:    store i32 [[TMP78]], i32* [[CONV13]], align 4
+// CHECK5-NEXT:    [[TMP79:%.*]] = load i64, i64* [[A_CASTED12]], align 8
+// CHECK5-NEXT:    [[TMP80:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK5-NEXT:    [[CONV15:%.*]] = bitcast i64* [[AA_CASTED14]] to i16*
-// CHECK5-NEXT:    store i16 [[TMP72]], i16* [[CONV15]], align 2
-// CHECK5-NEXT:    [[TMP73:%.*]] = load i64, i64* [[AA_CASTED14]], align 8
-// CHECK5-NEXT:    [[TMP74:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP74]], 10
+// CHECK5-NEXT:    store i16 [[TMP80]], i16* [[CONV15]], align 2
+// CHECK5-NEXT:    [[TMP81:%.*]] = load i64, i64* [[AA_CASTED14]], align 8
+// CHECK5-NEXT:    [[TMP82:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK5-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP82]], 10
 // CHECK5-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK5:       omp_if.then:
-// CHECK5-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP71]], i64* [[TMP76]], align 8
-// CHECK5-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP71]], i64* [[TMP78]], align 8
-// CHECK5-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP79]], align 8
-// CHECK5-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 1
-// CHECK5-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP73]], i64* [[TMP81]], align 8
-// CHECK5-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 1
-// CHECK5-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP73]], i64* [[TMP83]], align 8
-// CHECK5-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 1
-// CHECK5-NEXT:    store i8* null, i8** [[TMP84]], align 8
-// CHECK5-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP79]], i64* [[TMP84]], align 8
+// CHECK5-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP79]], i64* [[TMP86]], align 8
+// CHECK5-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP87]], align 8
+// CHECK5-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 1
+// CHECK5-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP81]], i64* [[TMP89]], align 8
+// CHECK5-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 1
+// CHECK5-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP81]], i64* [[TMP91]], align 8
+// CHECK5-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS18]], i64 0, i64 1
+// CHECK5-NEXT:    store i8* null, i8** [[TMP92]], align 8
+// CHECK5-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS16]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS17]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK5-NEXT:    [[TMP87:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP85]], i8** [[TMP86]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP88:%.*]] = icmp ne i32 [[TMP87]], 0
-// CHECK5-NEXT:    br i1 [[TMP88]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
-// CHECK5:       omp_offload.failed20:
-// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR4]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
-// CHECK5:       omp_offload.cont21:
+// CHECK5-NEXT:    [[KERNEL_ARGS20:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP95]], align 4
+// CHECK5-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 2, i32* [[TMP96]], align 4
+// CHECK5-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP93]], i8*** [[TMP97]], align 8
+// CHECK5-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP94]], i8*** [[TMP98]], align 8
+// CHECK5-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP99]], align 8
+// CHECK5-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP100]], align 8
+// CHECK5-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP101]], align 8
+// CHECK5-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP102]], align 8
+// CHECK5-NEXT:    [[TMP103:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]])
+// CHECK5-NEXT:    [[TMP104:%.*]] = icmp ne i32 [[TMP103]], 0
+// CHECK5-NEXT:    br i1 [[TMP104]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
+// CHECK5:       omp_offload.failed21:
+// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP79]], i64 [[TMP81]]) #[[ATTR4]]
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
+// CHECK5:       omp_offload.cont22:
 // CHECK5-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK5:       omp_if.else:
-// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP71]], i64 [[TMP73]]) #[[ATTR4]]
+// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i64 [[TMP79]], i64 [[TMP81]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_IF_END]]
 // CHECK5:       omp_if.end:
-// CHECK5-NEXT:    [[TMP89:%.*]] = load i32, i32* [[A]], align 4
-// CHECK5-NEXT:    [[CONV23:%.*]] = bitcast i64* [[A_CASTED22]] to i32*
-// CHECK5-NEXT:    store i32 [[TMP89]], i32* [[CONV23]], align 4
-// CHECK5-NEXT:    [[TMP90:%.*]] = load i64, i64* [[A_CASTED22]], align 8
-// CHECK5-NEXT:    [[TMP91:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK5-NEXT:    [[CMP24:%.*]] = icmp sgt i32 [[TMP91]], 20
-// CHECK5-NEXT:    br i1 [[CMP24]], label [[OMP_IF_THEN25:%.*]], label [[OMP_IF_ELSE32:%.*]]
-// CHECK5:       omp_if.then25:
-// CHECK5-NEXT:    [[TMP92:%.*]] = mul nuw i64 [[TMP2]], 4
-// CHECK5-NEXT:    [[TMP93:%.*]] = mul nuw i64 5, [[TMP5]]
-// CHECK5-NEXT:    [[TMP94:%.*]] = mul nuw i64 [[TMP93]], 8
-// CHECK5-NEXT:    [[TMP95:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK5-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP95]], i8* align 8 bitcast ([9 x i64]* @.offload_sizes.10 to i8*), i64 72, i1 false)
-// CHECK5-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP90]], i64* [[TMP97]], align 8
-// CHECK5-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP90]], i64* [[TMP99]], align 8
-// CHECK5-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP100]], align 8
-// CHECK5-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 1
-// CHECK5-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
-// CHECK5-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 8
-// CHECK5-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 1
-// CHECK5-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
-// CHECK5-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 8
-// CHECK5-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 1
-// CHECK5-NEXT:    store i8* null, i8** [[TMP105]], align 8
-// CHECK5-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 2
-// CHECK5-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP2]], i64* [[TMP107]], align 8
-// CHECK5-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 2
-// CHECK5-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP2]], i64* [[TMP109]], align 8
-// CHECK5-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 2
-// CHECK5-NEXT:    store i8* null, i8** [[TMP110]], align 8
-// CHECK5-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 3
-// CHECK5-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to float**
-// CHECK5-NEXT:    store float* [[VLA]], float** [[TMP112]], align 8
-// CHECK5-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 3
-// CHECK5-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
-// CHECK5-NEXT:    store float* [[VLA]], float** [[TMP114]], align 8
-// CHECK5-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK5-NEXT:    store i64 [[TMP92]], i64* [[TMP115]], align 8
-// CHECK5-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 3
+// CHECK5-NEXT:    [[TMP105:%.*]] = load i32, i32* [[A]], align 4
+// CHECK5-NEXT:    [[CONV24:%.*]] = bitcast i64* [[A_CASTED23]] to i32*
+// CHECK5-NEXT:    store i32 [[TMP105]], i32* [[CONV24]], align 4
+// CHECK5-NEXT:    [[TMP106:%.*]] = load i64, i64* [[A_CASTED23]], align 8
+// CHECK5-NEXT:    [[TMP107:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK5-NEXT:    [[CMP25:%.*]] = icmp sgt i32 [[TMP107]], 20
+// CHECK5-NEXT:    br i1 [[CMP25]], label [[OMP_IF_THEN26:%.*]], label [[OMP_IF_ELSE34:%.*]]
+// CHECK5:       omp_if.then26:
+// CHECK5-NEXT:    [[TMP108:%.*]] = mul nuw i64 [[TMP2]], 4
+// CHECK5-NEXT:    [[TMP109:%.*]] = mul nuw i64 5, [[TMP5]]
+// CHECK5-NEXT:    [[TMP110:%.*]] = mul nuw i64 [[TMP109]], 8
+// CHECK5-NEXT:    [[TMP111:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK5-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP111]], i8* align 8 bitcast ([9 x i64]* @.offload_sizes.10 to i8*), i64 72, i1 false)
+// CHECK5-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP106]], i64* [[TMP113]], align 8
+// CHECK5-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP106]], i64* [[TMP115]], align 8
+// CHECK5-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 0
 // CHECK5-NEXT:    store i8* null, i8** [[TMP116]], align 8
-// CHECK5-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 4
-// CHECK5-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to [5 x [10 x double]]**
-// CHECK5-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP118]], align 8
-// CHECK5-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 4
-// CHECK5-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
-// CHECK5-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 8
-// CHECK5-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 4
+// CHECK5-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 1
+// CHECK5-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to [10 x float]**
+// CHECK5-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP118]], align 8
+// CHECK5-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 1
+// CHECK5-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [10 x float]**
+// CHECK5-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP120]], align 8
+// CHECK5-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 1
 // CHECK5-NEXT:    store i8* null, i8** [[TMP121]], align 8
-// CHECK5-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 5
+// CHECK5-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 2
 // CHECK5-NEXT:    [[TMP123:%.*]] = bitcast i8** [[TMP122]] to i64*
-// CHECK5-NEXT:    store i64 5, i64* [[TMP123]], align 8
-// CHECK5-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 5
+// CHECK5-NEXT:    store i64 [[TMP2]], i64* [[TMP123]], align 8
+// CHECK5-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 2
 // CHECK5-NEXT:    [[TMP125:%.*]] = bitcast i8** [[TMP124]] to i64*
-// CHECK5-NEXT:    store i64 5, i64* [[TMP125]], align 8
-// CHECK5-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 5
+// CHECK5-NEXT:    store i64 [[TMP2]], i64* [[TMP125]], align 8
+// CHECK5-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 2
 // CHECK5-NEXT:    store i8* null, i8** [[TMP126]], align 8
-// CHECK5-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 6
-// CHECK5-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP5]], i64* [[TMP128]], align 8
-// CHECK5-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 6
-// CHECK5-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP5]], i64* [[TMP130]], align 8
-// CHECK5-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 6
-// CHECK5-NEXT:    store i8* null, i8** [[TMP131]], align 8
-// CHECK5-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 7
-// CHECK5-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to double**
-// CHECK5-NEXT:    store double* [[VLA1]], double** [[TMP133]], align 8
-// CHECK5-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 7
-// CHECK5-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to double**
-// CHECK5-NEXT:    store double* [[VLA1]], double** [[TMP135]], align 8
-// CHECK5-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK5-NEXT:    store i64 [[TMP94]], i64* [[TMP136]], align 8
-// CHECK5-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 7
+// CHECK5-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 3
+// CHECK5-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to float**
+// CHECK5-NEXT:    store float* [[VLA]], float** [[TMP128]], align 8
+// CHECK5-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 3
+// CHECK5-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to float**
+// CHECK5-NEXT:    store float* [[VLA]], float** [[TMP130]], align 8
+// CHECK5-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK5-NEXT:    store i64 [[TMP108]], i64* [[TMP131]], align 8
+// CHECK5-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 3
+// CHECK5-NEXT:    store i8* null, i8** [[TMP132]], align 8
+// CHECK5-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 4
+// CHECK5-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to [5 x [10 x double]]**
+// CHECK5-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP134]], align 8
+// CHECK5-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 4
+// CHECK5-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to [5 x [10 x double]]**
+// CHECK5-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP136]], align 8
+// CHECK5-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 4
 // CHECK5-NEXT:    store i8* null, i8** [[TMP137]], align 8
-// CHECK5-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 8
-// CHECK5-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to %struct.TT**
-// CHECK5-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP139]], align 8
-// CHECK5-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 8
-// CHECK5-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to %struct.TT**
-// CHECK5-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP141]], align 8
-// CHECK5-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS28]], i64 0, i64 8
+// CHECK5-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 5
+// CHECK5-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i64*
+// CHECK5-NEXT:    store i64 5, i64* [[TMP139]], align 8
+// CHECK5-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 5
+// CHECK5-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i64*
+// CHECK5-NEXT:    store i64 5, i64* [[TMP141]], align 8
+// CHECK5-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 5
 // CHECK5-NEXT:    store i8* null, i8** [[TMP142]], align 8
-// CHECK5-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS26]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS27]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 6
+// CHECK5-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP5]], i64* [[TMP144]], align 8
+// CHECK5-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 6
+// CHECK5-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP5]], i64* [[TMP146]], align 8
+// CHECK5-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 6
+// CHECK5-NEXT:    store i8* null, i8** [[TMP147]], align 8
+// CHECK5-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 7
+// CHECK5-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to double**
+// CHECK5-NEXT:    store double* [[VLA1]], double** [[TMP149]], align 8
+// CHECK5-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 7
+// CHECK5-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to double**
+// CHECK5-NEXT:    store double* [[VLA1]], double** [[TMP151]], align 8
+// CHECK5-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK5-NEXT:    store i64 [[TMP110]], i64* [[TMP152]], align 8
+// CHECK5-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 7
+// CHECK5-NEXT:    store i8* null, i8** [[TMP153]], align 8
+// CHECK5-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 8
+// CHECK5-NEXT:    [[TMP155:%.*]] = bitcast i8** [[TMP154]] to %struct.TT**
+// CHECK5-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP155]], align 8
+// CHECK5-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 8
+// CHECK5-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to %struct.TT**
+// CHECK5-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP157]], align 8
+// CHECK5-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 8
+// CHECK5-NEXT:    store i8* null, i8** [[TMP158]], align 8
+// CHECK5-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK5-NEXT:    [[TMP146:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP143]], i8** [[TMP144]], i64* [[TMP145]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.11, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP147:%.*]] = icmp ne i32 [[TMP146]], 0
-// CHECK5-NEXT:    br i1 [[TMP147]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK5:       omp_offload.failed30:
-// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK5:       omp_offload.cont31:
-// CHECK5-NEXT:    br label [[OMP_IF_END33:%.*]]
-// CHECK5:       omp_if.else32:
-// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP90]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
-// CHECK5-NEXT:    br label [[OMP_IF_END33]]
-// CHECK5:       omp_if.end33:
-// CHECK5-NEXT:    [[TMP148:%.*]] = load i32, i32* [[A]], align 4
-// CHECK5-NEXT:    [[TMP149:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK5-NEXT:    call void @llvm.stackrestore(i8* [[TMP149]])
-// CHECK5-NEXT:    ret i32 [[TMP148]]
+// CHECK5-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP162]], align 4
+// CHECK5-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 9, i32* [[TMP163]], align 4
+// CHECK5-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP159]], i8*** [[TMP164]], align 8
+// CHECK5-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP160]], i8*** [[TMP165]], align 8
+// CHECK5-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* [[TMP161]], i64** [[TMP166]], align 8
+// CHECK5-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.11, i32 0, i32 0), i64** [[TMP167]], align 8
+// CHECK5-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP168]], align 8
+// CHECK5-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP169]], align 8
+// CHECK5-NEXT:    [[TMP170:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK5-NEXT:    [[TMP171:%.*]] = icmp ne i32 [[TMP170]], 0
+// CHECK5-NEXT:    br i1 [[TMP171]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK5:       omp_offload.failed32:
+// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP106]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK5:       omp_offload.cont33:
+// CHECK5-NEXT:    br label [[OMP_IF_END35:%.*]]
+// CHECK5:       omp_if.else34:
+// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i64 [[TMP106]], [10 x float]* [[B]], i64 [[TMP2]], float* [[VLA]], [5 x [10 x double]]* [[C]], i64 5, i64 [[TMP5]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
+// CHECK5-NEXT:    br label [[OMP_IF_END35]]
+// CHECK5:       omp_if.end35:
+// CHECK5-NEXT:    [[TMP172:%.*]] = load i32, i32* [[A]], align 4
+// CHECK5-NEXT:    [[TMP173:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK5-NEXT:    call void @llvm.stackrestore(i8* [[TMP173]])
+// CHECK5-NEXT:    ret i32 [[TMP172]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97
@@ -4151,6 +4440,7 @@ int bar(int n){
 // CHECK5-NEXT:    [[AA_CASTED_I:%.*]] = alloca i64, align 8
 // CHECK5-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i64, align 8
 // CHECK5-NEXT:    [[DOTCAPTURE_EXPR__CASTED5_I:%.*]] = alloca i64, align 8
+// CHECK5-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK5-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
 // CHECK5-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -4191,23 +4481,39 @@ int bar(int n){
 // CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
 // CHECK5-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10) #[[ATTR4]]
-// CHECK5-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 1, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK5-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK5-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP26]], align 4, !noalias !26
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 3, i32* [[TMP27]], align 4, !noalias !26
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP20]], i8*** [[TMP28]], align 8, !noalias !26
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP21]], i8*** [[TMP29]], align 8, !noalias !26
+// CHECK5-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* [[TMP22]], i64** [[TMP30]], align 8, !noalias !26
+// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP31]], align 8, !noalias !26
+// CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP32]], align 8, !noalias !26
+// CHECK5-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP33]], align 8, !noalias !26
+// CHECK5-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP25]], i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK5-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK5-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK5:       omp_offload.failed.i:
-// CHECK5-NEXT:    [[TMP28:%.*]] = load i16, i16* [[TMP16]], align 2
+// CHECK5-NEXT:    [[TMP36:%.*]] = load i16, i16* [[TMP16]], align 2
 // CHECK5-NEXT:    [[CONV_I:%.*]] = bitcast i64* [[AA_CASTED_I]] to i16*
-// CHECK5-NEXT:    store i16 [[TMP28]], i16* [[CONV_I]], align 2, !noalias !26
-// CHECK5-NEXT:    [[TMP29:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !26
-// CHECK5-NEXT:    [[TMP30:%.*]] = load i32, i32* [[TMP23]], align 4
+// CHECK5-NEXT:    store i16 [[TMP36]], i16* [[CONV_I]], align 2, !noalias !26
+// CHECK5-NEXT:    [[TMP37:%.*]] = load i64, i64* [[AA_CASTED_I]], align 8, !noalias !26
+// CHECK5-NEXT:    [[TMP38:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK5-NEXT:    [[CONV4_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED_I]] to i32*
-// CHECK5-NEXT:    store i32 [[TMP30]], i32* [[CONV4_I]], align 4, !noalias !26
-// CHECK5-NEXT:    [[TMP31:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED_I]], align 8, !noalias !26
-// CHECK5-NEXT:    [[TMP32:%.*]] = load i32, i32* [[TMP24]], align 4
+// CHECK5-NEXT:    store i32 [[TMP38]], i32* [[CONV4_I]], align 4, !noalias !26
+// CHECK5-NEXT:    [[TMP39:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED_I]], align 8, !noalias !26
+// CHECK5-NEXT:    [[TMP40:%.*]] = load i32, i32* [[TMP24]], align 4
 // CHECK5-NEXT:    [[CONV6_I:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED5_I]] to i32*
-// CHECK5-NEXT:    store i32 [[TMP32]], i32* [[CONV6_I]], align 4, !noalias !26
-// CHECK5-NEXT:    [[TMP33:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5_I]], align 8, !noalias !26
-// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97(i64 [[TMP29]], i64 [[TMP31]], i64 [[TMP33]]) #[[ATTR4]]
+// CHECK5-NEXT:    store i32 [[TMP40]], i32* [[CONV6_I]], align 4, !noalias !26
+// CHECK5-NEXT:    [[TMP41:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5_I]], align 8, !noalias !26
+// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97(i64 [[TMP37]], i64 [[TMP39]], i64 [[TMP41]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
 // CHECK5:       .omp_outlined..1.exit:
 // CHECK5-NEXT:    ret i32 0
@@ -4804,9 +5110,26 @@ int bar(int n){
 // CHECK5-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [6 x i8*], [6 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [6 x i64], [6 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK5-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215.region_id, i32 6, i8** [[TMP45]], i8** [[TMP46]], i64* [[TMP47]], i64* getelementptr inbounds ([6 x i64], [6 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
-// CHECK5-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK5-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 6, i32* [[TMP49]], align 4
+// CHECK5-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP45]], i8*** [[TMP50]], align 8
+// CHECK5-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP46]], i8*** [[TMP51]], align 8
+// CHECK5-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* [[TMP47]], i64** [[TMP52]], align 8
+// CHECK5-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([6 x i64], [6 x i64]* @.offload_maptypes.14, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK5-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK5-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK5-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK5-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215(%struct.S1* [[THIS1]], i64 [[TMP7]], i64 2, i64 [[TMP2]], i16* [[VLA]], i64 [[TMP9]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -4816,15 +5139,15 @@ int bar(int n){
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215(%struct.S1* [[THIS1]], i64 [[TMP7]], i64 2, i64 [[TMP2]], i16* [[VLA]], i64 [[TMP9]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_IF_END]]
 // CHECK5:       omp_if.end:
-// CHECK5-NEXT:    [[TMP50:%.*]] = mul nsw i64 1, [[TMP2]]
-// CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP50]]
+// CHECK5-NEXT:    [[TMP58:%.*]] = mul nsw i64 1, [[TMP2]]
+// CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i64 [[TMP58]]
 // CHECK5-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i64 1
-// CHECK5-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX5]], align 2
-// CHECK5-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP51]] to i32
-// CHECK5-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
-// CHECK5-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], [[TMP52]]
-// CHECK5-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK5-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
+// CHECK5-NEXT:    [[TMP59:%.*]] = load i16, i16* [[ARRAYIDX5]], align 2
+// CHECK5-NEXT:    [[CONV6:%.*]] = sext i16 [[TMP59]] to i32
+// CHECK5-NEXT:    [[TMP60:%.*]] = load i32, i32* [[B]], align 4
+// CHECK5-NEXT:    [[ADD7:%.*]] = add nsw i32 [[CONV6]], [[TMP60]]
+// CHECK5-NEXT:    [[TMP61:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK5-NEXT:    call void @llvm.stackrestore(i8* [[TMP61]])
 // CHECK5-NEXT:    ret i32 [[ADD7]]
 //
 //
@@ -4929,9 +5252,26 @@ int bar(int n){
 // CHECK5-NEXT:    [[ADD8:%.*]] = add i32 [[TMP40]], 1
 // CHECK5-NEXT:    [[TMP41:%.*]] = zext i32 [[ADD8]] to i64
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP41]])
-// CHECK5-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK5-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK5-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 5, i32* [[TMP43]], align 4
+// CHECK5-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP34]], i8*** [[TMP44]], align 8
+// CHECK5-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP35]], i8*** [[TMP45]], align 8
+// CHECK5-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK5-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP47]], align 8
+// CHECK5-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK5-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP49]], align 8
+// CHECK5-NEXT:    [[TMP50:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP51:%.*]] = icmp ne i32 [[TMP50]], 0
+// CHECK5-NEXT:    br i1 [[TMP51]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], i64 [[TMP7]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -4941,8 +5281,8 @@ int bar(int n){
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197(i64 [[TMP1]], i64 [[TMP3]], i64 [[TMP5]], i64 [[TMP7]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_IF_END]]
 // CHECK5:       omp_if.end:
-// CHECK5-NEXT:    [[TMP44:%.*]] = load i32, i32* [[A]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP44]]
+// CHECK5-NEXT:    [[TMP52:%.*]] = load i32, i32* [[A]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP52]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -5000,9 +5340,26 @@ int bar(int n){
 // CHECK5-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK5-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.19, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.20, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK5-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK5-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 8
+// CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.19, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.20, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK5-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -5012,8 +5369,8 @@ int bar(int n){
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK5-NEXT:    br label [[OMP_IF_END]]
 // CHECK5:       omp_if.end:
-// CHECK5-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP24]]
+// CHECK5-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215
@@ -5530,12 +5887,12 @@ int bar(int n){
 // CHECK7-NEXT:    [[DOTOFFLOAD_PTRS13:%.*]] = alloca [2 x i8*], align 4
 // CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS14:%.*]] = alloca [2 x i8*], align 4
 // CHECK7-NEXT:    [[_TMP15:%.*]] = alloca i32, align 4
-// CHECK7-NEXT:    [[A_CASTED18:%.*]] = alloca i32, align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [9 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [9 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [9 x i8*], align 4
+// CHECK7-NEXT:    [[A_CASTED19:%.*]] = alloca i32, align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [9 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [9 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [9 x i8*], align 4
 // CHECK7-NEXT:    [[DOTOFFLOAD_SIZES:%.*]] = alloca [9 x i64], align 4
-// CHECK7-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
+// CHECK7-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
 // CHECK7-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
 // CHECK7-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK7-NEXT:    store i32 0, i32* [[A]], align 4
@@ -5641,165 +5998,216 @@ int bar(int n){
 // CHECK7-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS6]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS7]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK7-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, i32 1, i8** [[TMP64]], i8** [[TMP65]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK7-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
-// CHECK7-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK7-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK7-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK7-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP65]], i8*** [[TMP69]], align 4
+// CHECK7-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK7-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP71]], align 4
+// CHECK7-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK7-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP73]], align 4
+// CHECK7-NEXT:    [[TMP74:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP75:%.*]] = icmp ne i32 [[TMP74]], 0
+// CHECK7-NEXT:    br i1 [[TMP75]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l111(i32 [[TMP58]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK7:       omp_offload.cont:
-// CHECK7-NEXT:    [[TMP68:%.*]] = load i32, i32* [[A]], align 4
-// CHECK7-NEXT:    store i32 [[TMP68]], i32* [[A_CASTED9]], align 4
-// CHECK7-NEXT:    [[TMP69:%.*]] = load i32, i32* [[A_CASTED9]], align 4
-// CHECK7-NEXT:    [[TMP70:%.*]] = load i16, i16* [[AA]], align 2
+// CHECK7-NEXT:    [[TMP76:%.*]] = load i32, i32* [[A]], align 4
+// CHECK7-NEXT:    store i32 [[TMP76]], i32* [[A_CASTED9]], align 4
+// CHECK7-NEXT:    [[TMP77:%.*]] = load i32, i32* [[A_CASTED9]], align 4
+// CHECK7-NEXT:    [[TMP78:%.*]] = load i16, i16* [[AA]], align 2
 // CHECK7-NEXT:    [[CONV11:%.*]] = bitcast i32* [[AA_CASTED10]] to i16*
-// CHECK7-NEXT:    store i16 [[TMP70]], i16* [[CONV11]], align 2
-// CHECK7-NEXT:    [[TMP71:%.*]] = load i32, i32* [[AA_CASTED10]], align 4
-// CHECK7-NEXT:    [[TMP72:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK7-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP72]], 10
+// CHECK7-NEXT:    store i16 [[TMP78]], i16* [[CONV11]], align 2
+// CHECK7-NEXT:    [[TMP79:%.*]] = load i32, i32* [[AA_CASTED10]], align 4
+// CHECK7-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK7-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP80]], 10
 // CHECK7-NEXT:    br i1 [[CMP]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_ELSE:%.*]]
 // CHECK7:       omp_if.then:
-// CHECK7-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP69]], i32* [[TMP74]], align 4
-// CHECK7-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP69]], i32* [[TMP76]], align 4
-// CHECK7-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP77]], align 4
-// CHECK7-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
-// CHECK7-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP71]], i32* [[TMP79]], align 4
-// CHECK7-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
-// CHECK7-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP71]], i32* [[TMP81]], align 4
-// CHECK7-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 1
-// CHECK7-NEXT:    store i8* null, i8** [[TMP82]], align 4
-// CHECK7-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP77]], i32* [[TMP82]], align 4
+// CHECK7-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP77]], i32* [[TMP84]], align 4
+// CHECK7-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP85]], align 4
+// CHECK7-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
+// CHECK7-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP79]], i32* [[TMP87]], align 4
+// CHECK7-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
+// CHECK7-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP79]], i32* [[TMP89]], align 4
+// CHECK7-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i32 0, i32 1
+// CHECK7-NEXT:    store i8* null, i8** [[TMP90]], align 4
+// CHECK7-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK7-NEXT:    [[TMP85:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, i32 2, i8** [[TMP83]], i8** [[TMP84]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK7-NEXT:    [[TMP86:%.*]] = icmp ne i32 [[TMP85]], 0
-// CHECK7-NEXT:    br i1 [[TMP86]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK7:       omp_offload.failed16:
-// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR4]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK7:       omp_offload.cont17:
+// CHECK7-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP93]], align 4
+// CHECK7-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 2, i32* [[TMP94]], align 4
+// CHECK7-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP91]], i8*** [[TMP95]], align 4
+// CHECK7-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP92]], i8*** [[TMP96]], align 4
+// CHECK7-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP97]], align 4
+// CHECK7-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP98]], align 4
+// CHECK7-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP99]], align 4
+// CHECK7-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP100]], align 4
+// CHECK7-NEXT:    [[TMP101:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK7-NEXT:    [[TMP102:%.*]] = icmp ne i32 [[TMP101]], 0
+// CHECK7-NEXT:    br i1 [[TMP102]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK7:       omp_offload.failed17:
+// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP77]], i32 [[TMP79]]) #[[ATTR4]]
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK7:       omp_offload.cont18:
 // CHECK7-NEXT:    br label [[OMP_IF_END:%.*]]
 // CHECK7:       omp_if.else:
-// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP69]], i32 [[TMP71]]) #[[ATTR4]]
+// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l118(i32 [[TMP77]], i32 [[TMP79]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_IF_END]]
 // CHECK7:       omp_if.end:
-// CHECK7-NEXT:    [[TMP87:%.*]] = load i32, i32* [[A]], align 4
-// CHECK7-NEXT:    store i32 [[TMP87]], i32* [[A_CASTED18]], align 4
-// CHECK7-NEXT:    [[TMP88:%.*]] = load i32, i32* [[A_CASTED18]], align 4
-// CHECK7-NEXT:    [[TMP89:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK7-NEXT:    [[CMP19:%.*]] = icmp sgt i32 [[TMP89]], 20
-// CHECK7-NEXT:    br i1 [[CMP19]], label [[OMP_IF_THEN20:%.*]], label [[OMP_IF_ELSE27:%.*]]
-// CHECK7:       omp_if.then20:
-// CHECK7-NEXT:    [[TMP90:%.*]] = mul nuw i32 [[TMP1]], 4
-// CHECK7-NEXT:    [[TMP91:%.*]] = sext i32 [[TMP90]] to i64
-// CHECK7-NEXT:    [[TMP92:%.*]] = mul nuw i32 5, [[TMP3]]
-// CHECK7-NEXT:    [[TMP93:%.*]] = mul nuw i32 [[TMP92]], 8
-// CHECK7-NEXT:    [[TMP94:%.*]] = sext i32 [[TMP93]] to i64
-// CHECK7-NEXT:    [[TMP95:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
-// CHECK7-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP95]], i8* align 4 bitcast ([9 x i64]* @.offload_sizes.10 to i8*), i32 72, i1 false)
-// CHECK7-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP88]], i32* [[TMP97]], align 4
-// CHECK7-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP88]], i32* [[TMP99]], align 4
-// CHECK7-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP100]], align 4
-// CHECK7-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK7-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to [10 x float]**
-// CHECK7-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP102]], align 4
-// CHECK7-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK7-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to [10 x float]**
-// CHECK7-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP104]], align 4
-// CHECK7-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 1
-// CHECK7-NEXT:    store i8* null, i8** [[TMP105]], align 4
-// CHECK7-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
-// CHECK7-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP1]], i32* [[TMP107]], align 4
-// CHECK7-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
-// CHECK7-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP1]], i32* [[TMP109]], align 4
-// CHECK7-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 2
-// CHECK7-NEXT:    store i8* null, i8** [[TMP110]], align 4
-// CHECK7-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
-// CHECK7-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to float**
-// CHECK7-NEXT:    store float* [[VLA]], float** [[TMP112]], align 4
-// CHECK7-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
-// CHECK7-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to float**
-// CHECK7-NEXT:    store float* [[VLA]], float** [[TMP114]], align 4
-// CHECK7-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
-// CHECK7-NEXT:    store i64 [[TMP91]], i64* [[TMP115]], align 4
-// CHECK7-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 3
+// CHECK7-NEXT:    [[TMP103:%.*]] = load i32, i32* [[A]], align 4
+// CHECK7-NEXT:    store i32 [[TMP103]], i32* [[A_CASTED19]], align 4
+// CHECK7-NEXT:    [[TMP104:%.*]] = load i32, i32* [[A_CASTED19]], align 4
+// CHECK7-NEXT:    [[TMP105:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK7-NEXT:    [[CMP20:%.*]] = icmp sgt i32 [[TMP105]], 20
+// CHECK7-NEXT:    br i1 [[CMP20]], label [[OMP_IF_THEN21:%.*]], label [[OMP_IF_ELSE29:%.*]]
+// CHECK7:       omp_if.then21:
+// CHECK7-NEXT:    [[TMP106:%.*]] = mul nuw i32 [[TMP1]], 4
+// CHECK7-NEXT:    [[TMP107:%.*]] = sext i32 [[TMP106]] to i64
+// CHECK7-NEXT:    [[TMP108:%.*]] = mul nuw i32 5, [[TMP3]]
+// CHECK7-NEXT:    [[TMP109:%.*]] = mul nuw i32 [[TMP108]], 8
+// CHECK7-NEXT:    [[TMP110:%.*]] = sext i32 [[TMP109]] to i64
+// CHECK7-NEXT:    [[TMP111:%.*]] = bitcast [9 x i64]* [[DOTOFFLOAD_SIZES]] to i8*
+// CHECK7-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP111]], i8* align 4 bitcast ([9 x i64]* @.offload_sizes.10 to i8*), i32 72, i1 false)
+// CHECK7-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP104]], i32* [[TMP113]], align 4
+// CHECK7-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP104]], i32* [[TMP115]], align 4
+// CHECK7-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 0
 // CHECK7-NEXT:    store i8* null, i8** [[TMP116]], align 4
-// CHECK7-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 4
-// CHECK7-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to [5 x [10 x double]]**
-// CHECK7-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP118]], align 4
-// CHECK7-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 4
-// CHECK7-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [5 x [10 x double]]**
-// CHECK7-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP120]], align 4
-// CHECK7-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 4
+// CHECK7-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
+// CHECK7-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to [10 x float]**
+// CHECK7-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP118]], align 4
+// CHECK7-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
+// CHECK7-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to [10 x float]**
+// CHECK7-NEXT:    store [10 x float]* [[B]], [10 x float]** [[TMP120]], align 4
+// CHECK7-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 1
 // CHECK7-NEXT:    store i8* null, i8** [[TMP121]], align 4
-// CHECK7-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 5
+// CHECK7-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
 // CHECK7-NEXT:    [[TMP123:%.*]] = bitcast i8** [[TMP122]] to i32*
-// CHECK7-NEXT:    store i32 5, i32* [[TMP123]], align 4
-// CHECK7-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 5
+// CHECK7-NEXT:    store i32 [[TMP1]], i32* [[TMP123]], align 4
+// CHECK7-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
 // CHECK7-NEXT:    [[TMP125:%.*]] = bitcast i8** [[TMP124]] to i32*
-// CHECK7-NEXT:    store i32 5, i32* [[TMP125]], align 4
-// CHECK7-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 5
+// CHECK7-NEXT:    store i32 [[TMP1]], i32* [[TMP125]], align 4
+// CHECK7-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 2
 // CHECK7-NEXT:    store i8* null, i8** [[TMP126]], align 4
-// CHECK7-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 6
-// CHECK7-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP3]], i32* [[TMP128]], align 4
-// CHECK7-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 6
-// CHECK7-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to i32*
-// CHECK7-NEXT:    store i32 [[TMP3]], i32* [[TMP130]], align 4
-// CHECK7-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 6
-// CHECK7-NEXT:    store i8* null, i8** [[TMP131]], align 4
-// CHECK7-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 7
-// CHECK7-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to double**
-// CHECK7-NEXT:    store double* [[VLA1]], double** [[TMP133]], align 4
-// CHECK7-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 7
-// CHECK7-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to double**
-// CHECK7-NEXT:    store double* [[VLA1]], double** [[TMP135]], align 4
-// CHECK7-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
-// CHECK7-NEXT:    store i64 [[TMP94]], i64* [[TMP136]], align 4
-// CHECK7-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 7
+// CHECK7-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK7-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to float**
+// CHECK7-NEXT:    store float* [[VLA]], float** [[TMP128]], align 4
+// CHECK7-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK7-NEXT:    [[TMP130:%.*]] = bitcast i8** [[TMP129]] to float**
+// CHECK7-NEXT:    store float* [[VLA]], float** [[TMP130]], align 4
+// CHECK7-NEXT:    [[TMP131:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 3
+// CHECK7-NEXT:    store i64 [[TMP107]], i64* [[TMP131]], align 4
+// CHECK7-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 3
+// CHECK7-NEXT:    store i8* null, i8** [[TMP132]], align 4
+// CHECK7-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 4
+// CHECK7-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to [5 x [10 x double]]**
+// CHECK7-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP134]], align 4
+// CHECK7-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 4
+// CHECK7-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to [5 x [10 x double]]**
+// CHECK7-NEXT:    store [5 x [10 x double]]* [[C]], [5 x [10 x double]]** [[TMP136]], align 4
+// CHECK7-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 4
 // CHECK7-NEXT:    store i8* null, i8** [[TMP137]], align 4
-// CHECK7-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 8
-// CHECK7-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to %struct.TT**
-// CHECK7-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP139]], align 4
-// CHECK7-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 8
-// CHECK7-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to %struct.TT**
-// CHECK7-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP141]], align 4
-// CHECK7-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 8
+// CHECK7-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 5
+// CHECK7-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32*
+// CHECK7-NEXT:    store i32 5, i32* [[TMP139]], align 4
+// CHECK7-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 5
+// CHECK7-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32*
+// CHECK7-NEXT:    store i32 5, i32* [[TMP141]], align 4
+// CHECK7-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 5
 // CHECK7-NEXT:    store i8* null, i8** [[TMP142]], align 4
-// CHECK7-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 6
+// CHECK7-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP3]], i32* [[TMP144]], align 4
+// CHECK7-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 6
+// CHECK7-NEXT:    [[TMP146:%.*]] = bitcast i8** [[TMP145]] to i32*
+// CHECK7-NEXT:    store i32 [[TMP3]], i32* [[TMP146]], align 4
+// CHECK7-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 6
+// CHECK7-NEXT:    store i8* null, i8** [[TMP147]], align 4
+// CHECK7-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 7
+// CHECK7-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to double**
+// CHECK7-NEXT:    store double* [[VLA1]], double** [[TMP149]], align 4
+// CHECK7-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 7
+// CHECK7-NEXT:    [[TMP151:%.*]] = bitcast i8** [[TMP150]] to double**
+// CHECK7-NEXT:    store double* [[VLA1]], double** [[TMP151]], align 4
+// CHECK7-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 7
+// CHECK7-NEXT:    store i64 [[TMP110]], i64* [[TMP152]], align 4
+// CHECK7-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 7
+// CHECK7-NEXT:    store i8* null, i8** [[TMP153]], align 4
+// CHECK7-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 8
+// CHECK7-NEXT:    [[TMP155:%.*]] = bitcast i8** [[TMP154]] to %struct.TT**
+// CHECK7-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP155]], align 4
+// CHECK7-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 8
+// CHECK7-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to %struct.TT**
+// CHECK7-NEXT:    store %struct.TT* [[D]], %struct.TT** [[TMP157]], align 4
+// CHECK7-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i32 0, i32 8
+// CHECK7-NEXT:    store i8* null, i8** [[TMP158]], align 4
+// CHECK7-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP161:%.*]] = getelementptr inbounds [9 x i64], [9 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK7-NEXT:    [[TMP146:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, i32 9, i8** [[TMP143]], i8** [[TMP144]], i64* [[TMP145]], i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.11, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK7-NEXT:    [[TMP147:%.*]] = icmp ne i32 [[TMP146]], 0
-// CHECK7-NEXT:    br i1 [[TMP147]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
-// CHECK7:       omp_offload.failed25:
-// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
-// CHECK7:       omp_offload.cont26:
-// CHECK7-NEXT:    br label [[OMP_IF_END28:%.*]]
-// CHECK7:       omp_if.else27:
-// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP88]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
-// CHECK7-NEXT:    br label [[OMP_IF_END28]]
-// CHECK7:       omp_if.end28:
-// CHECK7-NEXT:    [[TMP148:%.*]] = load i32, i32* [[A]], align 4
-// CHECK7-NEXT:    [[TMP149:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK7-NEXT:    call void @llvm.stackrestore(i8* [[TMP149]])
-// CHECK7-NEXT:    ret i32 [[TMP148]]
+// CHECK7-NEXT:    [[KERNEL_ARGS26:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP162:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP162]], align 4
+// CHECK7-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 9, i32* [[TMP163]], align 4
+// CHECK7-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP159]], i8*** [[TMP164]], align 4
+// CHECK7-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP160]], i8*** [[TMP165]], align 4
+// CHECK7-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* [[TMP161]], i64** [[TMP166]], align 4
+// CHECK7-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([9 x i64], [9 x i64]* @.offload_maptypes.11, i32 0, i32 0), i64** [[TMP167]], align 4
+// CHECK7-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP168]], align 4
+// CHECK7-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP169]], align 4
+// CHECK7-NEXT:    [[TMP170:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS26]])
+// CHECK7-NEXT:    [[TMP171:%.*]] = icmp ne i32 [[TMP170]], 0
+// CHECK7-NEXT:    br i1 [[TMP171]], label [[OMP_OFFLOAD_FAILED27:%.*]], label [[OMP_OFFLOAD_CONT28:%.*]]
+// CHECK7:       omp_offload.failed27:
+// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP104]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT28]]
+// CHECK7:       omp_offload.cont28:
+// CHECK7-NEXT:    br label [[OMP_IF_END30:%.*]]
+// CHECK7:       omp_if.else29:
+// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l142(i32 [[TMP104]], [10 x float]* [[B]], i32 [[TMP1]], float* [[VLA]], [5 x [10 x double]]* [[C]], i32 5, i32 [[TMP3]], double* [[VLA1]], %struct.TT* [[D]]) #[[ATTR4]]
+// CHECK7-NEXT:    br label [[OMP_IF_END30]]
+// CHECK7:       omp_if.end30:
+// CHECK7-NEXT:    [[TMP172:%.*]] = load i32, i32* [[A]], align 4
+// CHECK7-NEXT:    [[TMP173:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK7-NEXT:    call void @llvm.stackrestore(i8* [[TMP173]])
+// CHECK7-NEXT:    ret i32 [[TMP172]]
 //
 //
 // CHECK7-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97
@@ -5940,6 +6348,7 @@ int bar(int n){
 // CHECK7-NEXT:    [[AA_CASTED_I:%.*]] = alloca i32, align 4
 // CHECK7-NEXT:    [[DOTCAPTURE_EXPR__CASTED_I:%.*]] = alloca i32, align 4
 // CHECK7-NEXT:    [[DOTCAPTURE_EXPR__CASTED4_I:%.*]] = alloca i32, align 4
+// CHECK7-NEXT:    [[KERNEL_ARGS_I:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
 // CHECK7-NEXT:    [[DOTADDR:%.*]] = alloca i32, align 4
 // CHECK7-NEXT:    [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 4
 // CHECK7-NEXT:    store i32 [[TMP0]], i32* [[DOTADDR]], align 4
@@ -5980,21 +6389,37 @@ int bar(int n){
 // CHECK7-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP12]], i32 0, i32 2
 // CHECK7-NEXT:    [[TMP25:%.*]] = load i32, i32* [[TMP23]], align 4
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10) #[[ATTR4]]
-// CHECK7-NEXT:    [[TMP26:%.*]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP25]], i32 1, i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
-// CHECK7-NEXT:    [[TMP27:%.*]] = icmp ne i32 [[TMP26]], 0
-// CHECK7-NEXT:    br i1 [[TMP27]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
+// CHECK7-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP26]], align 4, !noalias !27
+// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 3, i32* [[TMP27]], align 4, !noalias !27
+// CHECK7-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP20]], i8*** [[TMP28]], align 4, !noalias !27
+// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP21]], i8*** [[TMP29]], align 4, !noalias !27
+// CHECK7-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* [[TMP22]], i64** [[TMP30]], align 4, !noalias !27
+// CHECK7-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP31]], align 4, !noalias !27
+// CHECK7-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP32]], align 4, !noalias !27
+// CHECK7-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP33]], align 4, !noalias !27
+// CHECK7-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP25]], i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS_I]], i32 0, i8* null, i32 0, i8* null) #[[ATTR4]]
+// CHECK7-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
+// CHECK7-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED_I:%.*]], label [[DOTOMP_OUTLINED__1_EXIT:%.*]]
 // CHECK7:       omp_offload.failed.i:
-// CHECK7-NEXT:    [[TMP28:%.*]] = load i16, i16* [[TMP16]], align 2
+// CHECK7-NEXT:    [[TMP36:%.*]] = load i16, i16* [[TMP16]], align 2
 // CHECK7-NEXT:    [[CONV_I:%.*]] = bitcast i32* [[AA_CASTED_I]] to i16*
-// CHECK7-NEXT:    store i16 [[TMP28]], i16* [[CONV_I]], align 2, !noalias !27
-// CHECK7-NEXT:    [[TMP29:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !27
-// CHECK7-NEXT:    [[TMP30:%.*]] = load i32, i32* [[TMP23]], align 4
-// CHECK7-NEXT:    store i32 [[TMP30]], i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !27
-// CHECK7-NEXT:    [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !27
-// CHECK7-NEXT:    [[TMP32:%.*]] = load i32, i32* [[TMP24]], align 4
-// CHECK7-NEXT:    store i32 [[TMP32]], i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !27
-// CHECK7-NEXT:    [[TMP33:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !27
-// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97(i32 [[TMP29]], i32 [[TMP31]], i32 [[TMP33]]) #[[ATTR4]]
+// CHECK7-NEXT:    store i16 [[TMP36]], i16* [[CONV_I]], align 2, !noalias !27
+// CHECK7-NEXT:    [[TMP37:%.*]] = load i32, i32* [[AA_CASTED_I]], align 4, !noalias !27
+// CHECK7-NEXT:    [[TMP38:%.*]] = load i32, i32* [[TMP23]], align 4
+// CHECK7-NEXT:    store i32 [[TMP38]], i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !27
+// CHECK7-NEXT:    [[TMP39:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED_I]], align 4, !noalias !27
+// CHECK7-NEXT:    [[TMP40:%.*]] = load i32, i32* [[TMP24]], align 4
+// CHECK7-NEXT:    store i32 [[TMP40]], i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !27
+// CHECK7-NEXT:    [[TMP41:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4_I]], align 4, !noalias !27
+// CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z3fooi_l97(i32 [[TMP37]], i32 [[TMP39]], i32 [[TMP41]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[DOTOMP_OUTLINED__1_EXIT]]
 // CHECK7:       .omp_outlined..1.exit:
 // CHECK7-NEXT:    ret i32 0
@@ -6581,9 +7006,26 @@ int bar(int n){
 // CHECK7-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [6 x i8*], [6 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [6 x i64], [6 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK7-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215.region_id, i32 6, i8** [[TMP45]], i8** [[TMP46]], i64* [[TMP47]], i64* getelementptr inbounds ([6 x i64], [6 x i64]* @.offload_maptypes.14, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK7-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
-// CHECK7-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK7-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 6, i32* [[TMP49]], align 4
+// CHECK7-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP45]], i8*** [[TMP50]], align 4
+// CHECK7-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP46]], i8*** [[TMP51]], align 4
+// CHECK7-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* [[TMP47]], i64** [[TMP52]], align 4
+// CHECK7-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([6 x i64], [6 x i64]* @.offload_maptypes.14, i32 0, i32 0), i64** [[TMP53]], align 4
+// CHECK7-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP54]], align 4
+// CHECK7-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP55]], align 4
+// CHECK7-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK7-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215(%struct.S1* [[THIS1]], i32 [[TMP6]], i32 2, i32 [[TMP1]], i16* [[VLA]], i32 [[TMP8]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -6593,15 +7035,15 @@ int bar(int n){
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215(%struct.S1* [[THIS1]], i32 [[TMP6]], i32 2, i32 [[TMP1]], i16* [[VLA]], i32 [[TMP8]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_IF_END]]
 // CHECK7:       omp_if.end:
-// CHECK7-NEXT:    [[TMP50:%.*]] = mul nsw i32 1, [[TMP1]]
-// CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP50]]
+// CHECK7-NEXT:    [[TMP58:%.*]] = mul nsw i32 1, [[TMP1]]
+// CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[VLA]], i32 [[TMP58]]
 // CHECK7-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds i16, i16* [[ARRAYIDX]], i32 1
-// CHECK7-NEXT:    [[TMP51:%.*]] = load i16, i16* [[ARRAYIDX4]], align 2
-// CHECK7-NEXT:    [[CONV5:%.*]] = sext i16 [[TMP51]] to i32
-// CHECK7-NEXT:    [[TMP52:%.*]] = load i32, i32* [[B]], align 4
-// CHECK7-NEXT:    [[ADD6:%.*]] = add nsw i32 [[CONV5]], [[TMP52]]
-// CHECK7-NEXT:    [[TMP53:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK7-NEXT:    call void @llvm.stackrestore(i8* [[TMP53]])
+// CHECK7-NEXT:    [[TMP59:%.*]] = load i16, i16* [[ARRAYIDX4]], align 2
+// CHECK7-NEXT:    [[CONV5:%.*]] = sext i16 [[TMP59]] to i32
+// CHECK7-NEXT:    [[TMP60:%.*]] = load i32, i32* [[B]], align 4
+// CHECK7-NEXT:    [[ADD6:%.*]] = add nsw i32 [[CONV5]], [[TMP60]]
+// CHECK7-NEXT:    [[TMP61:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK7-NEXT:    call void @llvm.stackrestore(i8* [[TMP61]])
 // CHECK7-NEXT:    ret i32 [[ADD6]]
 //
 //
@@ -6704,9 +7146,26 @@ int bar(int n){
 // CHECK7-NEXT:    [[ADD6:%.*]] = add i32 [[TMP40]], 1
 // CHECK7-NEXT:    [[TMP41:%.*]] = zext i32 [[ADD6]] to i64
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP41]])
-// CHECK7-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK7-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
-// CHECK7-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK7-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 5, i32* [[TMP43]], align 4
+// CHECK7-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP34]], i8*** [[TMP44]], align 4
+// CHECK7-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP35]], i8*** [[TMP45]], align 4
+// CHECK7-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK7-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP47]], align 4
+// CHECK7-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK7-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP49]], align 4
+// CHECK7-NEXT:    [[TMP50:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP51:%.*]] = icmp ne i32 [[TMP50]], 0
+// CHECK7-NEXT:    br i1 [[TMP51]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], i32 [[TMP7]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -6716,8 +7175,8 @@ int bar(int n){
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l197(i32 [[TMP1]], i32 [[TMP3]], i32 [[TMP5]], i32 [[TMP7]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_IF_END]]
 // CHECK7:       omp_if.end:
-// CHECK7-NEXT:    [[TMP44:%.*]] = load i32, i32* [[A]], align 4
-// CHECK7-NEXT:    ret i32 [[TMP44]]
+// CHECK7-NEXT:    [[TMP52:%.*]] = load i32, i32* [[A]], align 4
+// CHECK7-NEXT:    ret i32 [[TMP52]]
 //
 //
 // CHECK7-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -6774,9 +7233,26 @@ int bar(int n){
 // CHECK7-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK7-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180.region_id, i32 3, i8** [[TMP20]], i8** [[TMP21]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.19, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.20, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK7-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK7-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK7-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK7-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP20]], i8*** [[TMP24]], align 4
+// CHECK7-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 4
+// CHECK7-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.19, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.20, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK7-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK7-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK7-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -6786,8 +7262,8 @@ int bar(int n){
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l180(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[B]]) #[[ATTR4]]
 // CHECK7-NEXT:    br label [[OMP_IF_END]]
 // CHECK7:       omp_if.end:
-// CHECK7-NEXT:    [[TMP24:%.*]] = load i32, i32* [[A]], align 4
-// CHECK7-NEXT:    ret i32 [[TMP24]]
+// CHECK7-NEXT:    [[TMP32:%.*]] = load i32, i32* [[A]], align 4
+// CHECK7-NEXT:    ret i32 [[TMP32]]
 //
 //
 // CHECK7-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l215

diff --git a/clang/test/OpenMP/target_teams_distribute_simd_collapse_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_simd_collapse_codegen.cpp
index 957b89d..da2057d 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_simd_collapse_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_simd_collapse_codegen.cpp
@@ -125,9 +125,26 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 56088)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR3:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -135,8 +152,8 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i64 0, i64 0
 // CHECK1-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP9]]
+// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -277,9 +294,26 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 56088)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR3:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -287,8 +321,8 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP9]]
+// CHECK3-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -648,20 +682,37 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP46:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_4]], align 8
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP46]], 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK9-NEXT:    [[TMP47:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80.region_id, i32 5, i8** [[TMP39]], i8** [[TMP40]], i64* [[TMP41]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP48:%.*]] = icmp ne i32 [[TMP47]], 0
-// CHECK9-NEXT:    br i1 [[TMP48]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP47]], align 4
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP48]], align 4
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 8
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP40]], i8*** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP41]], i64** [[TMP51]], align 8
+// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP56:%.*]] = icmp ne i32 [[TMP55]], 0
+// CHECK9-NEXT:    br i1 [[TMP56]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80(i64 [[TMP7]], i64 [[TMP9]], i64 [[TMP1]], i64 [[TMP3]], i32* [[VLA]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP49:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP49]])
+// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP57]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP50:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP50]])
-// CHECK9-NEXT:    [[TMP51:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP51]]
+// CHECK9-NEXT:    [[TMP58:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP58]])
+// CHECK9-NEXT:    [[TMP59:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP59]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80
@@ -884,9 +935,26 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 20)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l67.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l67.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l67([10 x [2 x i32]]* [[A]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1109,20 +1177,37 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP45:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_3]], align 8
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP45]], 1
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK11-NEXT:    [[TMP46:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80.region_id, i32 5, i8** [[TMP38]], i8** [[TMP39]], i64* [[TMP40]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP47:%.*]] = icmp ne i32 [[TMP46]], 0
-// CHECK11-NEXT:    br i1 [[TMP47]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP38]], i8*** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP40]], i64** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP55:%.*]] = icmp ne i32 [[TMP54]], 0
+// CHECK11-NEXT:    br i1 [[TMP55]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80(i32 [[TMP5]], i32 [[TMP7]], i32 [[TMP0]], i32 [[TMP1]], i32* [[VLA]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP48:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP48]])
+// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP56]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP49]])
-// CHECK11-NEXT:    [[TMP50:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP50]]
+// CHECK11-NEXT:    [[TMP57:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP57]])
+// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP58]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l80
@@ -1337,9 +1422,26 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 20)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l67.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l67.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l67([10 x [2 x i32]]* [[A]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/target_teams_distribute_simd_depend_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_simd_depend_codegen.cpp
index 64831d9..14ae36c 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_simd_depend_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_simd_depend_codegen.cpp
@@ -221,7 +221,7 @@ int foo(int n) {
 // CHECK:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 0
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
+// CHECK:       [[RET:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 0, i32 1, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 // CHECK-NEXT:  [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]
 // CHECK:       [[FAIL]]
@@ -242,8 +242,8 @@ int foo(int n) {
 // OMP50-32:       [[DEVICE_CAP:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* %{{.+}}, i32 0, i32 3
 // CHECK:       [[DEV:%.+]] = load i32, i32* [[DEVICE_CAP]],
 // CHECK:       [[DEVICE:%.+]] = sext i32 [[DEV]] to i64
-// OMP45:       [[RET:%.+]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 2, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SZT]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* [[MAPT]], i32 0, i32 0), i8** null, i8** null, i32 0, i32 1, i32 0, i8* null, i32 0, i8* null)
-// OMP50:       [[RET:%.+]] = call i32 @__tgt_target_teams_nowait_mapper(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i8* @{{[^,]+}}, i32 3, i8** [[BPR:%[^,]+]], i8** [[PR:%[^,]+]], i64* [[SZT]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* [[MAPT]], i32 0, i32 0), i8** null, i8** null, i32 0, i32 1, i32 0, i8* null, i32 0, i8* null)
+// OMP45:       [[RET:%.+]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 0, i32 1, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
+// OMP50:       [[RET:%.+]] = call i32 @__tgt_target_kernel_nowait(%struct.ident_t* @{{.+}}, i64 [[DEVICE]], i32 0, i32 1, i8* @.{{.+}}.region_id, %struct.__tgt_kernel_arguments* %{{.+}})
 
 // CHECK:       [[ERROR:%.+]] = icmp ne i32 [[RET]], 0
 // CHECK-NEXT:  br i1 [[ERROR]], label %[[FAIL:[^,]+]], label %[[END:[^,]+]]

diff --git a/clang/test/OpenMP/target_teams_distribute_simd_dist_schedule_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_simd_dist_schedule_codegen.cpp
index 42f815c..ab92189 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_simd_dist_schedule_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_simd_dist_schedule_codegen.cpp
@@ -142,10 +142,10 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -160,55 +160,106 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR3:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32(%struct.SS* [[THIS1]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK1:       omp_offload.failed16:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK1:       omp_offload.cont15:
-// CHECK1-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP27]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK1:       omp_offload.cont17:
+// CHECK1-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i64 0, i64 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -510,10 +561,10 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -528,55 +579,106 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR3:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK3:       omp_offload.failed8:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l32(%struct.SS* [[THIS1]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK3:       omp_offload.cont8:
-// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK3:       omp_offload.cont9:
+// CHECK3-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK3-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK3:       omp_offload.failed14:
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK3:       omp_offload.cont15:
-// CHECK3-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP27]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -1127,16 +1229,16 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[_TMP9:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
-// CHECK9-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES26:%.*]] = alloca [4 x i64], align 8
+// CHECK9-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_29:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -1194,142 +1296,193 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK9-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK9-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK9-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i64 24, i1 false)
-// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK9-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i64 24, i1 false)
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK9-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK9-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK9-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK9-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK9-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK9-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP58]])
-// CHECK9-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK9-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK9:       omp_offload.failed16:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR4]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK9:       omp_offload.cont17:
-// CHECK9-NEXT:    [[TMP61:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP61]], i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK9-NEXT:    [[TMP62:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP62]], i32* [[CONV20]], align 4
-// CHECK9-NEXT:    [[TMP63:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK9-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
-// CHECK9-NEXT:    [[CONV21:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP64]], i32* [[CONV21]], align 4
-// CHECK9-NEXT:    [[TMP65:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK9-NEXT:    [[TMP66:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP67:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP67]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.5 to i8*), i64 32, i1 false)
-// CHECK9-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP63]], i64* [[TMP69]], align 8
-// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP63]], i64* [[TMP71]], align 8
-// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP72]], align 8
-// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP74]], align 8
-// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP76]], align 8
-// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP77]], align 8
-// CHECK9-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP79]], align 8
-// CHECK9-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 8
-// CHECK9-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP66]], i64* [[TMP82]], align 8
-// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP83]], align 8
-// CHECK9-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK9-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP66]])
+// CHECK9-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK9-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK9-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK9-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK9-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK9-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK9-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK9:       omp_offload.failed17:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR4]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK9:       omp_offload.cont18:
+// CHECK9-NEXT:    [[TMP77:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP77]], i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK9-NEXT:    [[TMP78:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP78]], i32* [[CONV21]], align 4
+// CHECK9-NEXT:    [[TMP79:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK9-NEXT:    [[TMP80:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK9-NEXT:    [[CONV22:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP80]], i32* [[CONV22]], align 4
+// CHECK9-NEXT:    [[TMP81:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK9-NEXT:    [[TMP82:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP83:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES26]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP83]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.5 to i8*), i64 32, i1 false)
+// CHECK9-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP65]], i64* [[TMP85]], align 8
-// CHECK9-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK9-NEXT:    store i64 [[TMP79]], i64* [[TMP85]], align 8
+// CHECK9-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP65]], i64* [[TMP87]], align 8
-// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK9-NEXT:    store i64 [[TMP79]], i64* [[TMP87]], align 8
+// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 0
 // CHECK9-NEXT:    store i8* null, i8** [[TMP88]], align 8
-// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP92:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP92]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK9-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK9-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP93]], 0
-// CHECK9-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
-// CHECK9-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
-// CHECK9-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK9-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
-// CHECK9-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP94]], 1
-// CHECK9-NEXT:    [[TMP95:%.*]] = zext i32 [[ADD32]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP95]])
-// CHECK9-NEXT:    [[TMP96:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, i32 4, i8** [[TMP89]], i8** [[TMP90]], i64* [[TMP91]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP97:%.*]] = icmp ne i32 [[TMP96]], 0
-// CHECK9-NEXT:    br i1 [[TMP97]], label [[OMP_OFFLOAD_FAILED33:%.*]], label [[OMP_OFFLOAD_CONT34:%.*]]
-// CHECK9:       omp_offload.failed33:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i64 [[TMP63]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP65]]) #[[ATTR4]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT34]]
-// CHECK9:       omp_offload.cont34:
-// CHECK9-NEXT:    [[TMP98:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP98]])
+// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP90:%.*]] = bitcast i8** [[TMP89]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP90]], align 8
+// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP92]], align 8
+// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP93]], align 8
+// CHECK9-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP95:%.*]] = bitcast i8** [[TMP94]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP95]], align 8
+// CHECK9-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 8
+// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP82]], i64* [[TMP98]], align 8
+// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP99]], align 8
+// CHECK9-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP81]], i64* [[TMP101]], align 8
+// CHECK9-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP81]], i64* [[TMP103]], align 8
+// CHECK9-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP104]], align 8
+// CHECK9-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES26]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP108:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP108]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK9-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK9-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[TMP109]], 0
+// CHECK9-NEXT:    [[DIV31:%.*]] = sdiv i32 [[SUB30]], 1
+// CHECK9-NEXT:    [[SUB32:%.*]] = sub nsw i32 [[DIV31]], 1
+// CHECK9-NEXT:    store i32 [[SUB32]], i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK9-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_29]], align 4
+// CHECK9-NEXT:    [[ADD33:%.*]] = add nsw i32 [[TMP110]], 1
+// CHECK9-NEXT:    [[TMP111:%.*]] = zext i32 [[ADD33]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP111]])
+// CHECK9-NEXT:    [[KERNEL_ARGS34:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP112]], align 4
+// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP113]], align 4
+// CHECK9-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 8
+// CHECK9-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP107]], i64** [[TMP116]], align 8
+// CHECK9-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP117]], align 8
+// CHECK9-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK9-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP119]], align 8
+// CHECK9-NEXT:    [[TMP120:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS34]])
+// CHECK9-NEXT:    [[TMP121:%.*]] = icmp ne i32 [[TMP120]], 0
+// CHECK9-NEXT:    br i1 [[TMP121]], label [[OMP_OFFLOAD_FAILED35:%.*]], label [[OMP_OFFLOAD_CONT36:%.*]]
+// CHECK9:       omp_offload.failed35:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i64 [[TMP79]], i64 [[TMP1]], i32* [[VLA]], i64 [[TMP81]]) #[[ATTR4]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT36]]
+// CHECK9:       omp_offload.cont36:
+// CHECK9-NEXT:    [[TMP122:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP122]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP99:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP99]])
-// CHECK9-NEXT:    [[TMP100:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP100]]
+// CHECK9-NEXT:    [[TMP123:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP123]])
+// CHECK9-NEXT:    [[TMP124:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP124]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94
@@ -1758,10 +1911,10 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[DOTOFFLOAD_PTRS2:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [1 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [1 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [1 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP1:%.*]] = bitcast i8** [[TMP0]] to [10 x i32]**
@@ -1774,49 +1927,100 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l76.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l76.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l76([10 x i32]* [[A]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l80.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK9-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK9:       omp_offload.failed5:
+// CHECK9-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK9-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK9-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK9:       omp_offload.failed6:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l80([10 x i32]* [[A]]) #[[ATTR4]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK9:       omp_offload.cont6:
-// CHECK9-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP19]], align 8
-// CHECK9-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP21]], align 8
-// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK9:       omp_offload.cont7:
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK9-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK9:       omp_offload.failed11:
+// CHECK9-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK9-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK9-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK9:       omp_offload.failed13:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84([10 x i32]* [[A]]) #[[ATTR4]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK9:       omp_offload.cont12:
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK9:       omp_offload.cont14:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -2121,16 +2325,16 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[_TMP8:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_17:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_18:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED19:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR__CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
-// CHECK11-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES23:%.*]] = alloca [4 x i64], align 4
+// CHECK11-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -2187,141 +2391,192 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK11-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK11-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK11-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK11-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK11-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i32 24, i1 false)
-// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i32 24, i1 false)
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK11-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK11-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK11-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK11-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK11-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK11-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP59]])
-// CHECK11-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK11-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK11:       omp_offload.failed15:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR4]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK11:       omp_offload.cont16:
-// CHECK11-NEXT:    [[TMP62:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP62]], i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK11-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP63]], i32* [[N_CASTED18]], align 4
-// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N_CASTED18]], align 4
-// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK11-NEXT:    store i32 [[TMP65]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK11-NEXT:    [[TMP67:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP68:%.*]] = sext i32 [[TMP67]] to i64
-// CHECK11-NEXT:    [[TMP69:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP69]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.5 to i8*), i32 32, i1 false)
-// CHECK11-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[TMP71]], align 4
-// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[TMP73]], align 4
-// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP74]], align 4
-// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP76]], align 4
-// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP78]], align 4
-// CHECK11-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP79]], align 4
-// CHECK11-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP81]], align 4
-// CHECK11-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 4
-// CHECK11-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP68]], i64* [[TMP84]], align 4
-// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP85]], align 4
-// CHECK11-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK11-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP67]])
+// CHECK11-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK11-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK11-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK11-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK11-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK11-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK11-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK11:       omp_offload.failed16:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR4]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK11:       omp_offload.cont17:
+// CHECK11-NEXT:    [[TMP78:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP78]], i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK11-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP79]], i32* [[N_CASTED19]], align 4
+// CHECK11-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N_CASTED19]], align 4
+// CHECK11-NEXT:    [[TMP81:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK11-NEXT:    store i32 [[TMP81]], i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK11-NEXT:    [[TMP82:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK11-NEXT:    [[TMP83:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP84:%.*]] = sext i32 [[TMP83]] to i64
+// CHECK11-NEXT:    [[TMP85:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES23]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP85]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.5 to i8*), i32 32, i1 false)
+// CHECK11-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP66]], i32* [[TMP87]], align 4
-// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK11-NEXT:    store i32 [[TMP80]], i32* [[TMP87]], align 4
+// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP66]], i32* [[TMP89]], align 4
-// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK11-NEXT:    store i32 [[TMP80]], i32* [[TMP89]], align 4
+// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
 // CHECK11-NEXT:    store i8* null, i8** [[TMP90]], align 4
-// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP94:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP94]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP95]], 0
-// CHECK11-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
-// CHECK11-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
-// CHECK11-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK11-NEXT:    [[TMP96:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK11-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP96]], 1
-// CHECK11-NEXT:    [[TMP97:%.*]] = zext i32 [[ADD29]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP97]])
-// CHECK11-NEXT:    [[TMP98:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, i32 4, i8** [[TMP91]], i8** [[TMP92]], i64* [[TMP93]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP99:%.*]] = icmp ne i32 [[TMP98]], 0
-// CHECK11-NEXT:    br i1 [[TMP99]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
-// CHECK11:       omp_offload.failed30:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i32 [[TMP64]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP66]]) #[[ATTR4]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
-// CHECK11:       omp_offload.cont31:
-// CHECK11-NEXT:    [[TMP100:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP100]])
+// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP92]], align 4
+// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP94]], align 4
+// CHECK11-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP95]], align 4
+// CHECK11-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP97:%.*]] = bitcast i8** [[TMP96]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP97]], align 4
+// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 4
+// CHECK11-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP84]], i64* [[TMP100]], align 4
+// CHECK11-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP101]], align 4
+// CHECK11-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP82]], i32* [[TMP103]], align 4
+// CHECK11-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP82]], i32* [[TMP105]], align 4
+// CHECK11-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP106]], align 4
+// CHECK11-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP110:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP110]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP111]], 0
+// CHECK11-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
+// CHECK11-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
+// CHECK11-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK11-NEXT:    [[TMP112:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK11-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP112]], 1
+// CHECK11-NEXT:    [[TMP113:%.*]] = zext i32 [[ADD30]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP113]])
+// CHECK11-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP114]], align 4
+// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP115]], align 4
+// CHECK11-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK11-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP108]], i8*** [[TMP117]], align 4
+// CHECK11-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP109]], i64** [[TMP118]], align 4
+// CHECK11-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP119]], align 4
+// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK11-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP121]], align 4
+// CHECK11-NEXT:    [[TMP122:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK11-NEXT:    [[TMP123:%.*]] = icmp ne i32 [[TMP122]], 0
+// CHECK11-NEXT:    br i1 [[TMP123]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK11:       omp_offload.failed32:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l102(i32 [[TMP80]], i32 [[TMP0]], i32* [[VLA]], i32 [[TMP82]]) #[[ATTR4]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK11:       omp_offload.cont33:
+// CHECK11-NEXT:    [[TMP124:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP124]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP101:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP101]])
-// CHECK11-NEXT:    [[TMP102:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP102]]
+// CHECK11-NEXT:    [[TMP125:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP125]])
+// CHECK11-NEXT:    [[TMP126:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP126]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94
@@ -2735,10 +2990,10 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[DOTOFFLOAD_PTRS2:%.*]] = alloca [1 x i8*], align 4
 // CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 4
 // CHECK11-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [1 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [1 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [1 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [1 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [1 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [1 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP1:%.*]] = bitcast i8** [[TMP0]] to [10 x i32]**
@@ -2751,49 +3006,100 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l76.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l76.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l76([10 x i32]* [[A]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK11-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l80.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK11-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK11:       omp_offload.failed5:
+// CHECK11-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l80.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK11-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK11-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK11:       omp_offload.failed6:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l80([10 x i32]* [[A]]) #[[ATTR4]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK11:       omp_offload.cont6:
-// CHECK11-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP19]], align 4
-// CHECK11-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP21]], align 4
-// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK11:       omp_offload.cont7:
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK11-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK11:       omp_offload.failed11:
+// CHECK11-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK11-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK11-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK11:       omp_offload.failed13:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84([10 x i32]* [[A]]) #[[ATTR4]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK11:       omp_offload.cont12:
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK11:       omp_offload.cont14:
 // CHECK11-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/target_teams_distribute_simd_firstprivate_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_simd_firstprivate_codegen.cpp
index bd02ffc..a9d5b57 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_simd_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_simd_firstprivate_codegen.cpp
@@ -314,9 +314,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK1-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK1-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92([2 x i32]* @vec, i64 [[TMP1]], [2 x %struct.S]* @s_arr, %struct.S* @var, i64 [[TMP3]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -595,27 +612,44 @@ int main() {
 // CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK1-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK1-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56([2 x i32]* [[VEC]], i64 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP30]]
+// CHECK1-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN2StC2Ev
@@ -1116,9 +1150,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK3-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK3-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92([2 x i32]* @vec, i32 [[TMP1]], [2 x %struct.S]* @s_arr, %struct.S* @var, i32 [[TMP3]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1388,27 +1439,44 @@ int main() {
 // CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK3-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK3-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56([2 x i32]* [[VEC]], i32 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP30]]
+// CHECK3-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN2StC2Ev

diff --git a/clang/test/OpenMP/target_teams_distribute_simd_lastprivate_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_simd_lastprivate_codegen.cpp
index e7ac79a..cff5d27 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_simd_lastprivate_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_simd_lastprivate_codegen.cpp
@@ -613,9 +613,26 @@ int main() {
 // CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 5, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK9-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP35]], align 4
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP32]], i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP33]], i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK9-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94([2 x i32]* [[VEC]], i64 [[TMP3]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR5:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -623,18 +640,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP36]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP44]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR5]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done3:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR5]]
-// CHECK9-NEXT:    [[TMP37:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP37]]
+// CHECK9-NEXT:    [[TMP45:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP45]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -934,27 +951,44 @@ int main() {
 // CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK9-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK9-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49([2 x i32]* [[VEC]], i64 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR5]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR5]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR5]]
-// CHECK9-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP30]]
+// CHECK9-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1331,9 +1365,26 @@ int main() {
 // CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 5, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK11-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP32]], i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP33]], i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK11-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94([2 x i32]* [[VEC]], i32 [[TMP3]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR5:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1341,18 +1392,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP36]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP44]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR5]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR5]]
-// CHECK11-NEXT:    [[TMP37:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP37]]
+// CHECK11-NEXT:    [[TMP45:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP45]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1643,27 +1694,44 @@ int main() {
 // CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP25]], i8** [[TMP26]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK11-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP28]], align 4
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP25]], i8*** [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK11-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49([2 x i32]* [[VEC]], i32 [[TMP3]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR5]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP29]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP37]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR5]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR5]]
-// CHECK11-NEXT:    [[TMP30:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP30]]
+// CHECK11-NEXT:    [[TMP38:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP38]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/target_teams_distribute_simd_private_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_simd_private_codegen.cpp
index 2c2540f..f27481a 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_simd_private_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_simd_private_codegen.cpp
@@ -239,9 +239,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l91.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l91.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l91() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -392,27 +409,44 @@ int main() {
 // CHECK1-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 8
 // CHECK1-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 8
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK1-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP4]]
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev
@@ -741,9 +775,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l91.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l91.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l91() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -892,27 +943,44 @@ int main() {
 // CHECK3-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 4
 // CHECK3-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK3-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP4]]
+// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev

diff --git a/clang/test/OpenMP/target_teams_distribute_simd_reduction_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_simd_reduction_codegen.cpp
index b6f46ec..40e2f3e 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_simd_reduction_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_simd_reduction_codegen.cpp
@@ -98,9 +98,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l60.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l60.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l60(i32* @_ZZ4mainE5sivar) #[[ATTR3:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -265,9 +282,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
-// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
+// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i32* [[T_VAR]]) #[[ATTR3]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -435,9 +469,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l60.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l60.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l60(i32* @_ZZ4mainE5sivar) #[[ATTR3:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -602,9 +653,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
-// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
+// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i32* [[T_VAR]]) #[[ATTR3]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/target_teams_distribute_simd_uses_allocators_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_simd_uses_allocators_codegen.cpp
index b2fd8e0..88013d1 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_simd_uses_allocators_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_simd_uses_allocators_codegen.cpp
@@ -65,7 +65,7 @@ void foo() {
   omp_alloctrait_t traits[10];
   omp_allocator_handle_t my_allocator;
 
-// CHECK: [[RES:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @.[[TGT_REGION:.+]].region_id, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZES]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPTYPES]], i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
+// CHECK: [[RES:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 0, i32 1, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
 // CHECK: [[CMP:%.+]] = icmp ne i32 [[RES]], 0
 // CHECK: br i1 [[CMP]], label %[[FAILED:.+]], label %[[DONE:.+]]
 // CHECK: [[FAILED]]:

diff --git a/clang/test/OpenMP/target_teams_distribute_uses_allocators_codegen.cpp b/clang/test/OpenMP/target_teams_distribute_uses_allocators_codegen.cpp
index d6e1152..075e776 100644 (file)
--- a/clang/test/OpenMP/target_teams_distribute_uses_allocators_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_distribute_uses_allocators_codegen.cpp
@@ -65,7 +65,7 @@ void foo() {
   omp_alloctrait_t traits[10];
   omp_allocator_handle_t my_allocator;
 
-// CHECK: [[RES:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @.[[TGT_REGION:.+]].region_id, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZES]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPTYPES]], i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
+// CHECK: [[RES:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 0, i32 0, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
 // CHECK: [[CMP:%.+]] = icmp ne i32 [[RES]], 0
 // CHECK: br i1 [[CMP]], label %[[FAILED:.+]], label %[[DONE:.+]]
 // CHECK: [[FAILED]]:

diff --git a/clang/test/OpenMP/target_teams_map_codegen.cpp b/clang/test/OpenMP/target_teams_map_codegen.cpp
index c87a7be..f289e51 100644 (file)
--- a/clang/test/OpenMP/target_teams_map_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_map_codegen.cpp
@@ -102,9 +102,26 @@ void mapInt128() {
 // CHECK1-NEXT:    store i8* null, i8** [[TMP9]], align 8
 // CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14mapWithPrivatev_l27.region_id, i32 2, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
-// CHECK1-NEXT:    br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP13]], align 4
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP10]], i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP11]], i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14mapWithPrivatev_l27.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
+// CHECK1-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14mapWithPrivatev_l27() #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -157,9 +174,26 @@ void mapInt128() {
 // CHECK1-NEXT:    store i8* null, i8** [[TMP9]], align 8
 // CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z19mapWithFirstprivatev_l33.region_id, i32 2, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
-// CHECK1-NEXT:    br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP13]], align 4
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP10]], i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP11]], i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z19mapWithFirstprivatev_l33.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
+// CHECK1-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z19mapWithFirstprivatev_l33(i32* [[X]], i32* [[Y]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -232,9 +266,26 @@ void mapInt128() {
 // CHECK1-NEXT:    store i8* null, i8** [[TMP9]], align 8
 // CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16mapWithReductionv_l39.region_id, i32 2, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
-// CHECK1-NEXT:    br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP13]], align 4
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP10]], i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP11]], i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16mapWithReductionv_l39.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
+// CHECK1-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16mapWithReductionv_l39(i32* [[X]], i32* [[Y]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -360,9 +411,26 @@ void mapInt128() {
 // CHECK1-NEXT:    store i8* null, i8** [[TMP4]], align 8
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z7mapFromv_l45.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z7mapFromv_l45.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z7mapFromv_l45(i32* [[X]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -415,9 +483,26 @@ void mapInt128() {
 // CHECK1-NEXT:    store i8* null, i8** [[TMP4]], align 8
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5mapTov_l51.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5mapTov_l51.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5mapTov_l51(i32* [[X]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -470,9 +555,26 @@ void mapInt128() {
 // CHECK1-NEXT:    store i8* null, i8** [[TMP4]], align 8
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapAllocv_l57.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapAllocv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapAllocv_l57(i32* [[X]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -546,46 +648,80 @@ void mapInt128() {
 // CHECK1-NEXT:    store i8* null, i8** [[TMP14]], align 8
 // CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapArrayv_l63.region_id, i32 3, i8** [[TMP15]], i8** [[TMP16]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.18, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.19, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
-// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP17]], align 4
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP18]], align 4
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP15]], i8*** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP16]], i8*** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.18, i32 0, i32 0), i64** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.19, i32 0, i32 0), i64** [[TMP22]], align 8
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapArrayv_l63.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
+// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapArrayv_l63([88 x i32]* [[Y]], [99 x i32]* [[Z]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [88 x i32]**
-// CHECK1-NEXT:    store [88 x i32]* [[Y]], [88 x i32]** [[TMP20]], align 8
-// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [88 x i32]**
-// CHECK1-NEXT:    store [88 x i32]* [[Y]], [88 x i32]** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP23]], align 8
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to [99 x i32]**
-// CHECK1-NEXT:    store [99 x i32]* [[Z]], [99 x i32]** [[TMP25]], align 8
-// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to [99 x i32]**
-// CHECK1-NEXT:    store [99 x i32]* [[Z]], [99 x i32]** [[TMP27]], align 8
-// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP28]], align 8
-// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [77 x i32]**
-// CHECK1-NEXT:    store [77 x i32]* [[X]], [77 x i32]** [[TMP30]], align 8
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP32:%.*]] = bitcast i8** [[TMP31]] to [77 x i32]**
-// CHECK1-NEXT:    store [77 x i32]* [[X]], [77 x i32]** [[TMP32]], align 8
-// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP33]], align 8
-// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapArrayv_l65.region_id, i32 3, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.22, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.23, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK1-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED4:%.*]], label [[OMP_OFFLOAD_CONT5:%.*]]
-// CHECK1:       omp_offload.failed4:
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [88 x i32]**
+// CHECK1-NEXT:    store [88 x i32]* [[Y]], [88 x i32]** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [88 x i32]**
+// CHECK1-NEXT:    store [88 x i32]* [[Y]], [88 x i32]** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP33:%.*]] = bitcast i8** [[TMP32]] to [99 x i32]**
+// CHECK1-NEXT:    store [99 x i32]* [[Z]], [99 x i32]** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [99 x i32]**
+// CHECK1-NEXT:    store [99 x i32]* [[Z]], [99 x i32]** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [77 x i32]**
+// CHECK1-NEXT:    store [77 x i32]* [[X]], [77 x i32]** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [77 x i32]**
+// CHECK1-NEXT:    store [77 x i32]* [[X]], [77 x i32]** [[TMP40]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 2
+// CHECK1-NEXT:    store i8* null, i8** [[TMP41]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS4:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP44]], align 4
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP45]], align 4
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP42]], i8*** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP43]], i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.22, i32 0, i32 0), i64** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.23, i32 0, i32 0), i64** [[TMP49]], align 8
+// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP50]], align 8
+// CHECK1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP51]], align 8
+// CHECK1-NEXT:    [[TMP52:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapArrayv_l65.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]])
+// CHECK1-NEXT:    [[TMP53:%.*]] = icmp ne i32 [[TMP52]], 0
+// CHECK1-NEXT:    br i1 [[TMP53]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
+// CHECK1:       omp_offload.failed5:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapArrayv_l65([88 x i32]* [[Y]], [99 x i32]* [[Z]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT5]]
-// CHECK1:       omp_offload.cont5:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
+// CHECK1:       omp_offload.cont6:
 // CHECK1-NEXT:    ret void
 //
 //
@@ -887,46 +1023,80 @@ void mapInt128() {
 // CHECK1-NEXT:    store i8* null, i8** [[TMP14]], align 8
 // CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9mapInt128v_l72.region_id, i32 3, i8** [[TMP15]], i8** [[TMP16]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.26, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.27, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
-// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP17]], align 4
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP18]], align 4
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP15]], i8*** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP16]], i8*** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.26, i32 0, i32 0), i64** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.27, i32 0, i32 0), i64** [[TMP22]], align 8
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9mapInt128v_l72.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
+// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9mapInt128v_l72(i128* [[Y]], i128* [[Z]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i128**
-// CHECK1-NEXT:    store i128* [[Y]], i128** [[TMP20]], align 8
-// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to i128**
-// CHECK1-NEXT:    store i128* [[Y]], i128** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP23]], align 8
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i128**
-// CHECK1-NEXT:    store i128* [[Z]], i128** [[TMP25]], align 8
-// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i128**
-// CHECK1-NEXT:    store i128* [[Z]], i128** [[TMP27]], align 8
-// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP28]], align 8
-// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i128**
+// CHECK1-NEXT:    store i128* [[Y]], i128** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to i128**
-// CHECK1-NEXT:    store i128* [[X]], i128** [[TMP30]], align 8
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP32:%.*]] = bitcast i8** [[TMP31]] to i128**
-// CHECK1-NEXT:    store i128* [[X]], i128** [[TMP32]], align 8
-// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP33]], align 8
-// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9mapInt128v_l74.region_id, i32 3, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.30, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.31, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK1-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED4:%.*]], label [[OMP_OFFLOAD_CONT5:%.*]]
-// CHECK1:       omp_offload.failed4:
+// CHECK1-NEXT:    store i128* [[Y]], i128** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP33:%.*]] = bitcast i8** [[TMP32]] to i128**
+// CHECK1-NEXT:    store i128* [[Z]], i128** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i128**
+// CHECK1-NEXT:    store i128* [[Z]], i128** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i128**
+// CHECK1-NEXT:    store i128* [[X]], i128** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i128**
+// CHECK1-NEXT:    store i128* [[X]], i128** [[TMP40]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 2
+// CHECK1-NEXT:    store i8* null, i8** [[TMP41]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS4:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP44]], align 4
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP45]], align 4
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP42]], i8*** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP43]], i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.30, i32 0, i32 0), i64** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.31, i32 0, i32 0), i64** [[TMP49]], align 8
+// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP50]], align 8
+// CHECK1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP51]], align 8
+// CHECK1-NEXT:    [[TMP52:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9mapInt128v_l74.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]])
+// CHECK1-NEXT:    [[TMP53:%.*]] = icmp ne i32 [[TMP52]], 0
+// CHECK1-NEXT:    br i1 [[TMP53]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
+// CHECK1:       omp_offload.failed5:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9mapInt128v_l74(i128* [[Y]], i128* [[Z]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT5]]
-// CHECK1:       omp_offload.cont5:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
+// CHECK1:       omp_offload.cont6:
 // CHECK1-NEXT:    ret void
 //
 //
@@ -1169,9 +1339,26 @@ void mapInt128() {
 // CHECK3-NEXT:    store i8* null, i8** [[TMP9]], align 4
 // CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14mapWithPrivatev_l27.region_id, i32 2, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
-// CHECK3-NEXT:    br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP10]], i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP11]], i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP19]], align 4
+// CHECK3-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14mapWithPrivatev_l27.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
+// CHECK3-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z14mapWithPrivatev_l27() #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1224,9 +1411,26 @@ void mapInt128() {
 // CHECK3-NEXT:    store i8* null, i8** [[TMP9]], align 4
 // CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z19mapWithFirstprivatev_l33.region_id, i32 2, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
-// CHECK3-NEXT:    br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP10]], i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP11]], i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP19]], align 4
+// CHECK3-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z19mapWithFirstprivatev_l33.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
+// CHECK3-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z19mapWithFirstprivatev_l33(i32* [[X]], i32* [[Y]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1295,9 +1499,26 @@ void mapInt128() {
 // CHECK3-NEXT:    store i8* null, i8** [[TMP9]], align 4
 // CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16mapWithReductionv_l39.region_id, i32 2, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
-// CHECK3-NEXT:    br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP10]], i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP11]], i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP19]], align 4
+// CHECK3-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16mapWithReductionv_l39.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
+// CHECK3-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z16mapWithReductionv_l39(i32* [[X]], i32* [[Y]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1423,9 +1644,26 @@ void mapInt128() {
 // CHECK3-NEXT:    store i8* null, i8** [[TMP4]], align 4
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z7mapFromv_l45.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z7mapFromv_l45.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z7mapFromv_l45(i32* [[X]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1476,9 +1714,26 @@ void mapInt128() {
 // CHECK3-NEXT:    store i8* null, i8** [[TMP4]], align 4
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5mapTov_l51.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5mapTov_l51.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5mapTov_l51(i32* [[X]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1529,9 +1784,26 @@ void mapInt128() {
 // CHECK3-NEXT:    store i8* null, i8** [[TMP4]], align 4
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapAllocv_l57.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapAllocv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapAllocv_l57(i32* [[X]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1603,46 +1875,80 @@ void mapInt128() {
 // CHECK3-NEXT:    store i8* null, i8** [[TMP14]], align 4
 // CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapArrayv_l63.region_id, i32 3, i8** [[TMP15]], i8** [[TMP16]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.18, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.19, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
-// CHECK3-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP15]], i8*** [[TMP19]], align 4
+// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP16]], i8*** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.18, i32 0, i32 0), i64** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.19, i32 0, i32 0), i64** [[TMP22]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapArrayv_l63.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
+// CHECK3-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapArrayv_l63([88 x i32]* [[Y]], [99 x i32]* [[Z]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [88 x i32]**
-// CHECK3-NEXT:    store [88 x i32]* [[Y]], [88 x i32]** [[TMP20]], align 4
-// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to [88 x i32]**
-// CHECK3-NEXT:    store [88 x i32]* [[Y]], [88 x i32]** [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP23]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to [99 x i32]**
-// CHECK3-NEXT:    store [99 x i32]* [[Z]], [99 x i32]** [[TMP25]], align 4
-// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to [99 x i32]**
-// CHECK3-NEXT:    store [99 x i32]* [[Z]], [99 x i32]** [[TMP27]], align 4
-// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP28]], align 4
-// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [77 x i32]**
-// CHECK3-NEXT:    store [77 x i32]* [[X]], [77 x i32]** [[TMP30]], align 4
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP32:%.*]] = bitcast i8** [[TMP31]] to [77 x i32]**
-// CHECK3-NEXT:    store [77 x i32]* [[X]], [77 x i32]** [[TMP32]], align 4
-// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP33]], align 4
-// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapArrayv_l65.region_id, i32 3, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.22, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.23, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK3-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED4:%.*]], label [[OMP_OFFLOAD_CONT5:%.*]]
-// CHECK3:       omp_offload.failed4:
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [88 x i32]**
+// CHECK3-NEXT:    store [88 x i32]* [[Y]], [88 x i32]** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [88 x i32]**
+// CHECK3-NEXT:    store [88 x i32]* [[Y]], [88 x i32]** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP33:%.*]] = bitcast i8** [[TMP32]] to [99 x i32]**
+// CHECK3-NEXT:    store [99 x i32]* [[Z]], [99 x i32]** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [99 x i32]**
+// CHECK3-NEXT:    store [99 x i32]* [[Z]], [99 x i32]** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to [77 x i32]**
+// CHECK3-NEXT:    store [77 x i32]* [[X]], [77 x i32]** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to [77 x i32]**
+// CHECK3-NEXT:    store [77 x i32]* [[X]], [77 x i32]** [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 2
+// CHECK3-NEXT:    store i8* null, i8** [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS4:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP42]], i8*** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP43]], i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.22, i32 0, i32 0), i64** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.23, i32 0, i32 0), i64** [[TMP49]], align 4
+// CHECK3-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP50]], align 4
+// CHECK3-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP51]], align 4
+// CHECK3-NEXT:    [[TMP52:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapArrayv_l65.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]])
+// CHECK3-NEXT:    [[TMP53:%.*]] = icmp ne i32 [[TMP52]], 0
+// CHECK3-NEXT:    br i1 [[TMP53]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
+// CHECK3:       omp_offload.failed5:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z8mapArrayv_l65([88 x i32]* [[Y]], [99 x i32]* [[Z]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT5]]
-// CHECK3:       omp_offload.cont5:
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
+// CHECK3:       omp_offload.cont6:
 // CHECK3-NEXT:    ret void
 //
 //

diff --git a/clang/test/OpenMP/target_teams_num_teams_codegen.cpp b/clang/test/OpenMP/target_teams_num_teams_codegen.cpp
index 8ecefe2..7b353aa 100644 (file)
--- a/clang/test/OpenMP/target_teams_num_teams_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_num_teams_codegen.cpp
@@ -260,35 +260,69 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK1-NEXT:    [[TMP24:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, i32 3, i8** [[TMP21]], i8** [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP23]], i32 0)
-// CHECK1-NEXT:    [[TMP25:%.*]] = icmp ne i32 [[TMP24]], 0
-// CHECK1-NEXT:    br i1 [[TMP25]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 [[TMP23]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121(%struct.S1* [[THIS1]], i64 [[TMP3]], i64 [[TMP5]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to %struct.S1**
-// CHECK1-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP27]], align 8
-// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to double**
-// CHECK1-NEXT:    store double* [[A3]], double** [[TMP29]], align 8
-// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP30]], align 8
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, i32 1, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 1024, i32 0)
-// CHECK1-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK1-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.S1**
+// CHECK1-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to double**
+// CHECK1-NEXT:    store double* [[A3]], double** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1024, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP35:%.*]] = load double, double* [[A9]], align 8
-// CHECK1-NEXT:    [[CONV10:%.*]] = fptosi double [[TMP35]] to i32
-// CHECK1-NEXT:    ret i32 [[CONV10]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = load double, double* [[A10]], align 8
+// CHECK1-NEXT:    [[CONV11:%.*]] = fptosi double [[TMP51]] to i32
+// CHECK1-NEXT:    ret i32 [[CONV11]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZL7fstatici
@@ -323,41 +357,75 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK1-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 [[TMP10]], i32 0)
-// CHECK1-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK1-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP9]], i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP10]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP20:%.*]] = icmp ne i32 [[TMP19]], 0
+// CHECK1-NEXT:    br i1 [[TMP20]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104(i64 [[TMP2]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP13:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 32, [[TMP13]]
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 32, [[TMP21]]
 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK1-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK1-NEXT:    [[TMP22:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
 // CHECK1-NEXT:    [[CONV3:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED2]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP14]], i32* [[CONV3]], align 4
-// CHECK1-NEXT:    [[TMP15:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED2]], align 8
-// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP15]], i64* [[TMP17]], align 8
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP15]], i64* [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP20]], align 8
-// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK1-NEXT:    [[TMP24:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, i32 1, i8** [[TMP21]], i8** [[TMP22]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 [[TMP23]], i32 0)
-// CHECK1-NEXT:    [[TMP25:%.*]] = icmp ne i32 [[TMP24]], 0
-// CHECK1-NEXT:    br i1 [[TMP25]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108(i64 [[TMP15]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[TMP26:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP26]], 1
-// CHECK1-NEXT:    ret i32 [[ADD9]]
+// CHECK1-NEXT:    store i32 [[TMP22]], i32* [[CONV3]], align 4
+// CHECK1-NEXT:    [[TMP23:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED2]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP23]], i64* [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP23]], i64* [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP29]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP30]], i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP31]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK1-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108(i64 [[TMP23]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[TMP42:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[ADD10:%.*]] = add nsw i32 [[TMP42]], 1
+// CHECK1-NEXT:    ret i32 [[ADD10]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -375,65 +443,99 @@ int bar(int n){
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 8
 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK1-NEXT:    store i32 0, i32* [[A]], align 4
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 20, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 20, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i16 1, i16* [[B]], align 2
-// CHECK1-NEXT:    [[TMP2:%.*]] = load i16, i16* [[B]], align 2
-// CHECK1-NEXT:    store i16 [[TMP2]], i16* [[DOTCAPTURE_EXPR_]], align 2
-// CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = load i16, i16* [[B]], align 2
+// CHECK1-NEXT:    store i16 [[TMP10]], i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK1-NEXT:    [[TMP11:%.*]] = load i32, i32* [[A]], align 4
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP3]], i32* [[CONV]], align 4
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i16, i16* [[B]], align 2
+// CHECK1-NEXT:    store i32 [[TMP11]], i32* [[CONV]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = load i16, i16* [[B]], align 2
 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[B_CASTED]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP5]], i16* [[CONV1]], align 2
-// CHECK1-NEXT:    [[TMP6:%.*]] = load i64, i64* [[B_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP7:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK1-NEXT:    store i16 [[TMP13]], i16* [[CONV1]], align 2
+// CHECK1-NEXT:    [[TMP14:%.*]] = load i64, i64* [[B_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
 // CHECK1-NEXT:    [[CONV2:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP7]], i16* [[CONV2]], align 2
-// CHECK1-NEXT:    [[TMP8:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP15]], align 8
-// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP17]], align 8
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP18]], align 8
-// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK1-NEXT:    store i16 [[TMP15]], i16* [[CONV2]], align 2
+// CHECK1-NEXT:    [[TMP16:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP8]], i64* [[TMP20]], align 8
-// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP8]], i64* [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP23]], align 8
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP26:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
-// CHECK1-NEXT:    [[TMP27:%.*]] = sext i16 [[TMP26]] to i32
-// CHECK1-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 [[TMP27]], i32 0)
-// CHECK1-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
-// CHECK1-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
-// CHECK1:       omp_offload.failed3:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i64 [[TMP4]], i64 [[TMP6]], i64 [[TMP8]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
-// CHECK1:       omp_offload.cont4:
-// CHECK1-NEXT:    [[TMP30:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP30]]
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP16]], i64* [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP16]], i64* [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
+// CHECK1-NEXT:    store i8* null, i8** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK1-NEXT:    [[TMP35:%.*]] = sext i16 [[TMP34]] to i32
+// CHECK1-NEXT:    [[KERNEL_ARGS3:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP37]], align 4
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP32]], i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP33]], i8*** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP35]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]])
+// CHECK1-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK1-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED4:%.*]], label [[OMP_OFFLOAD_CONT5:%.*]]
+// CHECK1:       omp_offload.failed4:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i64 [[TMP12]], i64 [[TMP14]], i64 [[TMP16]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT5]]
+// CHECK1:       omp_offload.cont5:
+// CHECK1-NEXT:    [[TMP46:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP46]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121
@@ -715,34 +817,68 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, i32 3, i8** [[TMP21]], i8** [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP23]], i32 0)
-// CHECK3-NEXT:    [[TMP25:%.*]] = icmp ne i32 [[TMP24]], 0
-// CHECK3-NEXT:    br i1 [[TMP25]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 [[TMP23]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121(%struct.S1* [[THIS1]], i32 [[TMP3]], i32 [[TMP5]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to %struct.S1**
-// CHECK3-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP27]], align 4
-// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to double**
-// CHECK3-NEXT:    store double* [[A2]], double** [[TMP29]], align 4
-// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP30]], align 4
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, i32 1, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 1024, i32 0)
-// CHECK3-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK3-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK3:       omp_offload.failed6:
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.S1**
+// CHECK3-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to double**
+// CHECK3-NEXT:    store double* [[A2]], double** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 1024, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK3:       omp_offload.failed7:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK3:       omp_offload.cont7:
-// CHECK3-NEXT:    [[A8:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP35:%.*]] = load double, double* [[A8]], align 4
-// CHECK3-NEXT:    [[CONV:%.*]] = fptosi double [[TMP35]] to i32
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK3:       omp_offload.cont8:
+// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = load double, double* [[A9]], align 4
+// CHECK3-NEXT:    [[CONV:%.*]] = fptosi double [[TMP51]] to i32
 // CHECK3-NEXT:    ret i32 [[CONV]]
 //
 //
@@ -777,40 +913,74 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 [[TMP10]], i32 0)
-// CHECK3-NEXT:    [[TMP12:%.*]] = icmp ne i32 [[TMP11]], 0
-// CHECK3-NEXT:    br i1 [[TMP12]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP9]], i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP10]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP20:%.*]] = icmp ne i32 [[TMP19]], 0
+// CHECK3-NEXT:    br i1 [[TMP20]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104(i32 [[TMP2]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP13:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 32, [[TMP13]]
+// CHECK3-NEXT:    [[TMP21:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 32, [[TMP21]]
 // CHECK3-NEXT:    store i32 [[ADD]], i32* [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK3-NEXT:    store i32 [[TMP14]], i32* [[DOTCAPTURE_EXPR__CASTED2]], align 4
-// CHECK3-NEXT:    [[TMP15:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED2]], align 4
-// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP15]], i32* [[TMP17]], align 4
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP15]], i32* [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP20]], align 4
-// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, i32 1, i8** [[TMP21]], i8** [[TMP22]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 [[TMP23]], i32 0)
-// CHECK3-NEXT:    [[TMP25:%.*]] = icmp ne i32 [[TMP24]], 0
-// CHECK3-NEXT:    br i1 [[TMP25]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK3:       omp_offload.failed6:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108(i32 [[TMP15]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK3:       omp_offload.cont7:
-// CHECK3-NEXT:    [[TMP26:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[ADD8:%.*]] = add nsw i32 [[TMP26]], 1
-// CHECK3-NEXT:    ret i32 [[ADD8]]
+// CHECK3-NEXT:    [[TMP22:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK3-NEXT:    store i32 [[TMP22]], i32* [[DOTCAPTURE_EXPR__CASTED2]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED2]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP23]], i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP23]], i32* [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK3-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP29]], i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP30]], i8*** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP39]], align 4
+// CHECK3-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP31]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK3-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK3-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108(i32 [[TMP23]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK3:       omp_offload.cont8:
+// CHECK3-NEXT:    [[TMP42:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[ADD9:%.*]] = add nsw i32 [[TMP42]], 1
+// CHECK3-NEXT:    ret i32 [[ADD9]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -828,64 +998,98 @@ int bar(int n){
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 4
 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK3-NEXT:    store i32 0, i32* [[A]], align 4
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 20, i32 0)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 20, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i16 1, i16* [[B]], align 2
-// CHECK3-NEXT:    [[TMP2:%.*]] = load i16, i16* [[B]], align 2
-// CHECK3-NEXT:    store i16 [[TMP2]], i16* [[DOTCAPTURE_EXPR_]], align 2
-// CHECK3-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[A_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP5:%.*]] = load i16, i16* [[B]], align 2
+// CHECK3-NEXT:    [[TMP10:%.*]] = load i16, i16* [[B]], align 2
+// CHECK3-NEXT:    store i16 [[TMP10]], i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK3-NEXT:    [[TMP11:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP11]], i32* [[A_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[A_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = load i16, i16* [[B]], align 2
 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[B_CASTED]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP5]], i16* [[CONV]], align 2
-// CHECK3-NEXT:    [[TMP6:%.*]] = load i32, i32* [[B_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP7:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK3-NEXT:    store i16 [[TMP13]], i16* [[CONV]], align 2
+// CHECK3-NEXT:    [[TMP14:%.*]] = load i32, i32* [[B_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
 // CHECK3-NEXT:    [[CONV1:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR__CASTED]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP7]], i16* [[CONV1]], align 2
-// CHECK3-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP4]], i32* [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP4]], i32* [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP6]], i32* [[TMP15]], align 4
-// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP6]], i32* [[TMP17]], align 4
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP18]], align 4
-// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK3-NEXT:    store i16 [[TMP15]], i16* [[CONV1]], align 2
+// CHECK3-NEXT:    [[TMP16:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP8]], i32* [[TMP20]], align 4
-// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP8]], i32* [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP23]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP26:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
-// CHECK3-NEXT:    [[TMP27:%.*]] = sext i16 [[TMP26]] to i32
-// CHECK3-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 [[TMP27]], i32 0)
-// CHECK3-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
-// CHECK3-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK3:       omp_offload.failed2:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i32 [[TMP4]], i32 [[TMP6]], i32 [[TMP8]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK3:       omp_offload.cont3:
-// CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP30]]
+// CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP14]], i32* [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP14]], i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP16]], i32* [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP16]], i32* [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8* null, i8** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK3-NEXT:    [[TMP35:%.*]] = sext i16 [[TMP34]] to i32
+// CHECK3-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP32]], i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP33]], i8*** [[TMP39]], align 4
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP35]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK3-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK3-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK3:       omp_offload.failed3:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i32 [[TMP12]], i32 [[TMP14]], i32 [[TMP16]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK3:       omp_offload.cont4:
+// CHECK3-NEXT:    [[TMP46:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP46]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121

diff --git a/clang/test/OpenMP/target_teams_thread_limit_codegen.cpp b/clang/test/OpenMP/target_teams_thread_limit_codegen.cpp
index 2735f6f..e68a52e 100644 (file)
--- a/clang/test/OpenMP/target_teams_thread_limit_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_thread_limit_codegen.cpp
@@ -260,35 +260,69 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK1-NEXT:    [[TMP24:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, i32 3, i8** [[TMP21]], i8** [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP23]])
-// CHECK1-NEXT:    [[TMP25:%.*]] = icmp ne i32 [[TMP24]], 0
-// CHECK1-NEXT:    br i1 [[TMP25]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 [[TMP23]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121(%struct.S1* [[THIS1]], i64 [[TMP3]], i64 [[TMP5]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to %struct.S1**
-// CHECK1-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP27]], align 8
-// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to double**
-// CHECK1-NEXT:    store double* [[A3]], double** [[TMP29]], align 8
-// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP30]], align 8
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, i32 1, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1024)
-// CHECK1-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK1-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.S1**
+// CHECK1-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to double**
+// CHECK1-NEXT:    store double* [[A3]], double** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 1024, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP35:%.*]] = load double, double* [[A9]], align 8
-// CHECK1-NEXT:    [[CONV10:%.*]] = fptosi double [[TMP35]] to i32
-// CHECK1-NEXT:    ret i32 [[CONV10]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = load double, double* [[A10]], align 8
+// CHECK1-NEXT:    [[CONV11:%.*]] = fptosi double [[TMP51]] to i32
+// CHECK1-NEXT:    ret i32 [[CONV11]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZL7fstatici
@@ -341,41 +375,75 @@ int bar(int n){
 // CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP18:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
 // CHECK1-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK1-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, i32 2, i8** [[TMP16]], i8** [[TMP17]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 [[TMP18]], i32 [[TMP19]])
-// CHECK1-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
-// CHECK1-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP20]], align 4
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP21]], align 4
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP16]], i8*** [[TMP22]], align 8
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP17]], i8*** [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP18]], i32 [[TMP19]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
+// CHECK1-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104(i64 [[TMP3]], i64 [[TMP5]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP22:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 32, [[TMP22]]
+// CHECK1-NEXT:    [[TMP30:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 32, [[TMP30]]
 // CHECK1-NEXT:    store i32 [[ADD]], i32* [[DOTCAPTURE_EXPR_4]], align 4
-// CHECK1-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_4]], align 4
+// CHECK1-NEXT:    [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_4]], align 4
 // CHECK1-NEXT:    [[CONV6:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED5]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP23]], i32* [[CONV6]], align 4
-// CHECK1-NEXT:    [[TMP24:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5]], align 8
-// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP24]], i64* [[TMP26]], align 8
-// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP24]], i64* [[TMP28]], align 8
-// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP29]], align 8
-// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP32:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_4]], align 4
-// CHECK1-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, i32 1, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP32]])
-// CHECK1-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK1-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED10:%.*]], label [[OMP_OFFLOAD_CONT11:%.*]]
-// CHECK1:       omp_offload.failed10:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108(i64 [[TMP24]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT11]]
-// CHECK1:       omp_offload.cont11:
-// CHECK1-NEXT:    [[TMP35:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    [[ADD12:%.*]] = add nsw i32 [[TMP35]], 1
-// CHECK1-NEXT:    ret i32 [[ADD12]]
+// CHECK1-NEXT:    store i32 [[TMP31]], i32* [[CONV6]], align 4
+// CHECK1-NEXT:    [[TMP32:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED5]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = bitcast i8** [[TMP33]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP32]], i64* [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP32]], i64* [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_4]], align 4
+// CHECK1-NEXT:    [[KERNEL_ARGS10:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 [[TMP40]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
+// CHECK1:       omp_offload.failed11:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108(i64 [[TMP32]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
+// CHECK1:       omp_offload.cont12:
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[ADD13:%.*]] = add nsw i32 [[TMP51]], 1
+// CHECK1-NEXT:    ret i32 [[ADD13]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -393,65 +461,99 @@ int bar(int n){
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 8
 // CHECK1-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK1-NEXT:    store i32 0, i32* [[A]], align 4
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 20)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 20, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i16 1, i16* [[B]], align 2
-// CHECK1-NEXT:    [[TMP2:%.*]] = load i16, i16* [[B]], align 2
-// CHECK1-NEXT:    store i16 [[TMP2]], i16* [[DOTCAPTURE_EXPR_]], align 2
-// CHECK1-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = load i16, i16* [[B]], align 2
+// CHECK1-NEXT:    store i16 [[TMP10]], i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK1-NEXT:    [[TMP11:%.*]] = load i32, i32* [[A]], align 4
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP3]], i32* [[CONV]], align 4
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i16, i16* [[B]], align 2
+// CHECK1-NEXT:    store i32 [[TMP11]], i32* [[CONV]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = load i16, i16* [[B]], align 2
 // CHECK1-NEXT:    [[CONV1:%.*]] = bitcast i64* [[B_CASTED]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP5]], i16* [[CONV1]], align 2
-// CHECK1-NEXT:    [[TMP6:%.*]] = load i64, i64* [[B_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP7:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK1-NEXT:    store i16 [[TMP13]], i16* [[CONV1]], align 2
+// CHECK1-NEXT:    [[TMP14:%.*]] = load i64, i64* [[B_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
 // CHECK1-NEXT:    [[CONV2:%.*]] = bitcast i64* [[DOTCAPTURE_EXPR__CASTED]] to i16*
-// CHECK1-NEXT:    store i16 [[TMP7]], i16* [[CONV2]], align 2
-// CHECK1-NEXT:    [[TMP8:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP4]], i64* [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP15]], align 8
-// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP17]], align 8
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP18]], align 8
-// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK1-NEXT:    store i16 [[TMP15]], i16* [[CONV2]], align 2
+// CHECK1-NEXT:    [[TMP16:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR__CASTED]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP8]], i64* [[TMP20]], align 8
-// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP8]], i64* [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP23]], align 8
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP26:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
-// CHECK1-NEXT:    [[TMP27:%.*]] = sext i16 [[TMP26]] to i32
-// CHECK1-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 [[TMP27]], i32 1024)
-// CHECK1-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
-// CHECK1-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
-// CHECK1:       omp_offload.failed3:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i64 [[TMP4]], i64 [[TMP6]], i64 [[TMP8]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
-// CHECK1:       omp_offload.cont4:
-// CHECK1-NEXT:    [[TMP30:%.*]] = load i32, i32* [[A]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP30]]
+// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP16]], i64* [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP16]], i64* [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 2
+// CHECK1-NEXT:    store i8* null, i8** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK1-NEXT:    [[TMP35:%.*]] = sext i16 [[TMP34]] to i32
+// CHECK1-NEXT:    [[KERNEL_ARGS3:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP37]], align 4
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP32]], i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP33]], i8*** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP35]], i32 1024, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS3]])
+// CHECK1-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK1-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED4:%.*]], label [[OMP_OFFLOAD_CONT5:%.*]]
+// CHECK1:       omp_offload.failed4:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i64 [[TMP12]], i64 [[TMP14]], i64 [[TMP16]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT5]]
+// CHECK1:       omp_offload.cont5:
+// CHECK1-NEXT:    [[TMP46:%.*]] = load i32, i32* [[A]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP46]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121
@@ -737,34 +839,68 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, i32 3, i8** [[TMP21]], i8** [[TMP22]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP23]])
-// CHECK3-NEXT:    [[TMP25:%.*]] = icmp ne i32 [[TMP24]], 0
-// CHECK3-NEXT:    br i1 [[TMP25]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP21]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 [[TMP23]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121(%struct.S1* [[THIS1]], i32 [[TMP3]], i32 [[TMP5]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to %struct.S1**
-// CHECK3-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP27]], align 4
-// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to double**
-// CHECK3-NEXT:    store double* [[A2]], double** [[TMP29]], align 4
-// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP30]], align 4
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, i32 1, i8** [[TMP31]], i8** [[TMP32]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1024)
-// CHECK3-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK3-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK3:       omp_offload.failed6:
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.S1**
+// CHECK3-NEXT:    store %struct.S1* [[THIS1]], %struct.S1** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to double**
+// CHECK3-NEXT:    store double* [[A2]], double** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 1024, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK3:       omp_offload.failed7:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l126(%struct.S1* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK3:       omp_offload.cont7:
-// CHECK3-NEXT:    [[A8:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP35:%.*]] = load double, double* [[A8]], align 4
-// CHECK3-NEXT:    [[CONV:%.*]] = fptosi double [[TMP35]] to i32
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK3:       omp_offload.cont8:
+// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_S1]], %struct.S1* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = load double, double* [[A9]], align 4
+// CHECK3-NEXT:    [[CONV:%.*]] = fptosi double [[TMP51]] to i32
 // CHECK3-NEXT:    ret i32 [[CONV]]
 //
 //
@@ -816,40 +952,74 @@ int bar(int n){
 // CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP18:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4
 // CHECK3-NEXT:    [[TMP19:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, i32 2, i8** [[TMP16]], i8** [[TMP17]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 [[TMP18]], i32 [[TMP19]])
-// CHECK3-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
-// CHECK3-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP16]], i8*** [[TMP22]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP17]], i8*** [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP18]], i32 [[TMP19]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
+// CHECK3-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l104(i32 [[TMP3]], i32 [[TMP5]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP22:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 32, [[TMP22]]
+// CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 32, [[TMP30]]
 // CHECK3-NEXT:    store i32 [[ADD]], i32* [[DOTCAPTURE_EXPR_3]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_3]], align 4
-// CHECK3-NEXT:    store i32 [[TMP23]], i32* [[DOTCAPTURE_EXPR__CASTED4]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4]], align 4
-// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP24]], i32* [[TMP26]], align 4
-// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP24]], i32* [[TMP28]], align 4
-// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP29]], align 4
-// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_3]], align 4
-// CHECK3-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, i32 1, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP32]])
-// CHECK3-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
-// CHECK3-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
-// CHECK3:       omp_offload.failed8:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108(i32 [[TMP24]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
-// CHECK3:       omp_offload.cont9:
-// CHECK3-NEXT:    [[TMP35:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    [[ADD10:%.*]] = add nsw i32 [[TMP35]], 1
-// CHECK3-NEXT:    ret i32 [[ADD10]]
+// CHECK3-NEXT:    [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_3]], align 4
+// CHECK3-NEXT:    store i32 [[TMP31]], i32* [[DOTCAPTURE_EXPR__CASTED4]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED4]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = bitcast i8** [[TMP33]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP32]], i32* [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP32]], i32* [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_3]], align 4
+// CHECK3-NEXT:    [[KERNEL_ARGS8:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 [[TMP40]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED9:%.*]], label [[OMP_OFFLOAD_CONT10:%.*]]
+// CHECK3:       omp_offload.failed9:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZL7fstatici_l108(i32 [[TMP32]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT10]]
+// CHECK3:       omp_offload.cont10:
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    [[ADD11:%.*]] = add nsw i32 [[TMP51]], 1
+// CHECK3-NEXT:    ret i32 [[ADD11]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_Z9ftemplateIiET_i
@@ -867,64 +1037,98 @@ int bar(int n){
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [3 x i8*], align 4
 // CHECK3-NEXT:    store i32 [[N]], i32* [[N_ADDR]], align 4
 // CHECK3-NEXT:    store i32 0, i32* [[A]], align 4
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 20)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 20, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l88() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i16 1, i16* [[B]], align 2
-// CHECK3-NEXT:    [[TMP2:%.*]] = load i16, i16* [[B]], align 2
-// CHECK3-NEXT:    store i16 [[TMP2]], i16* [[DOTCAPTURE_EXPR_]], align 2
-// CHECK3-NEXT:    [[TMP3:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    store i32 [[TMP3]], i32* [[A_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP5:%.*]] = load i16, i16* [[B]], align 2
+// CHECK3-NEXT:    [[TMP10:%.*]] = load i16, i16* [[B]], align 2
+// CHECK3-NEXT:    store i16 [[TMP10]], i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK3-NEXT:    [[TMP11:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    store i32 [[TMP11]], i32* [[A_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[A_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = load i16, i16* [[B]], align 2
 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[B_CASTED]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP5]], i16* [[CONV]], align 2
-// CHECK3-NEXT:    [[TMP6:%.*]] = load i32, i32* [[B_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP7:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK3-NEXT:    store i16 [[TMP13]], i16* [[CONV]], align 2
+// CHECK3-NEXT:    [[TMP14:%.*]] = load i32, i32* [[B_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
 // CHECK3-NEXT:    [[CONV1:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR__CASTED]] to i16*
-// CHECK3-NEXT:    store i16 [[TMP7]], i16* [[CONV1]], align 2
-// CHECK3-NEXT:    [[TMP8:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP4]], i32* [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP4]], i32* [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP15:%.*]] = bitcast i8** [[TMP14]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP6]], i32* [[TMP15]], align 4
-// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP17:%.*]] = bitcast i8** [[TMP16]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP6]], i32* [[TMP17]], align 4
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP18]], align 4
-// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK3-NEXT:    store i16 [[TMP15]], i16* [[CONV1]], align 2
+// CHECK3-NEXT:    [[TMP16:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR__CASTED]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP8]], i32* [[TMP20]], align 4
-// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP8]], i32* [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP23]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP26:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
-// CHECK3-NEXT:    [[TMP27:%.*]] = sext i16 [[TMP26]] to i32
-// CHECK3-NEXT:    [[TMP28:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 [[TMP27]], i32 1024)
-// CHECK3-NEXT:    [[TMP29:%.*]] = icmp ne i32 [[TMP28]], 0
-// CHECK3-NEXT:    br i1 [[TMP29]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK3:       omp_offload.failed2:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i32 [[TMP4]], i32 [[TMP6]], i32 [[TMP8]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK3:       omp_offload.cont3:
-// CHECK3-NEXT:    [[TMP30:%.*]] = load i32, i32* [[A]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP30]]
+// CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP14]], i32* [[TMP23]], align 4
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP14]], i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP16]], i32* [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP16]], i32* [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8* null, i8** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = load i16, i16* [[DOTCAPTURE_EXPR_]], align 2
+// CHECK3-NEXT:    [[TMP35:%.*]] = sext i16 [[TMP34]] to i32
+// CHECK3-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP32]], i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP33]], i8*** [[TMP39]], align 4
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP35]], i32 1024, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK3-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK3-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK3:       omp_offload.failed3:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l93(i32 [[TMP12]], i32 [[TMP14]], i32 [[TMP16]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK3:       omp_offload.cont4:
+// CHECK3-NEXT:    [[TMP46:%.*]] = load i32, i32* [[A]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP46]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2S12r1Ei_l121

diff --git a/clang/test/OpenMP/target_teams_uses_allocators_codegen.cpp b/clang/test/OpenMP/target_teams_uses_allocators_codegen.cpp
index 568ac1d..644accc 100644 (file)
--- a/clang/test/OpenMP/target_teams_uses_allocators_codegen.cpp
+++ b/clang/test/OpenMP/target_teams_uses_allocators_codegen.cpp
@@ -65,7 +65,7 @@ void foo() {
   omp_alloctrait_t traits[10];
   omp_allocator_handle_t my_allocator;
 
-// CHECK: [[RES:%.+]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @.[[TGT_REGION:.+]].region_id, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZES]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPTYPES]], i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
+// CHECK: [[RES:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 0, i32 0, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
 // CHECK: [[CMP:%.+]] = icmp ne i32 [[RES]], 0
 // CHECK: br i1 [[CMP]], label %[[FAILED:.+]], label %[[DONE:.+]]
 // CHECK: [[FAILED]]:

diff --git a/clang/test/OpenMP/target_uses_allocators_codegen.cpp b/clang/test/OpenMP/target_uses_allocators_codegen.cpp
index 8cd1cc2..2a43ae8 100644 (file)
--- a/clang/test/OpenMP/target_uses_allocators_codegen.cpp
+++ b/clang/test/OpenMP/target_uses_allocators_codegen.cpp
@@ -65,7 +65,7 @@ void foo() {
   omp_alloctrait_t traits[10];
   omp_allocator_handle_t my_allocator;
 
-// CHECK: [[RES:%.+]] = call i32 @__tgt_target_mapper(%struct.ident_t* @{{.+}}, i64 -1, i8* @.[[TGT_REGION:.+]].region_id, i32 1, i8** %{{.+}}, i8** %{{.+}}, i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[SIZES]], i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* [[MAPTYPES]], i32 0, i32 0), i8** null, i8** null)
+// CHECK: [[RES:%.+]] = call i32 @__tgt_target_kernel(%struct.ident_t* @{{.+}}, i64 -1, i32 -1, i32 0, i8* @.[[TGT_REGION:.+]].region_id, %struct.__tgt_kernel_arguments* %[[KERNEL_ARGS:.+]])
 // CHECK: [[CMP:%.+]] = icmp ne i32 [[RES]], 0
 // CHECK: br i1 [[CMP]], label %[[FAILED:.+]], label %[[DONE:.+]]
 // CHECK: [[FAILED]]:

diff --git a/clang/test/OpenMP/teams_codegen.cpp b/clang/test/OpenMP/teams_codegen.cpp
index 772feac..8aab940 100644 (file)
--- a/clang/test/OpenMP/teams_codegen.cpp
+++ b/clang/test/OpenMP/teams_codegen.cpp
@@ -314,31 +314,31 @@ void foo() {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[LA_CASTED:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[COMP_CASTED9:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [2 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [2 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [2 x i8*], align 8
-// CHECK1-NEXT:    [[LA_CASTED16:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[COMP_CASTED18:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [2 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [2 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [2 x i8*], align 8
+// CHECK1-NEXT:    [[COMP_CASTED10:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS12:%.*]] = alloca [2 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS13:%.*]] = alloca [2 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS14:%.*]] = alloca [2 x i8*], align 8
+// CHECK1-NEXT:    [[LA_CASTED18:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[COMP_CASTED20:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [2 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [2 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [2 x i8*], align 8
 // CHECK1-NEXT:    [[GBLA_CASTED:%.*]] = alloca i64, align 8
 // CHECK1-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8
 // CHECK1-NEXT:    [[GBLB_CASTED:%.*]] = alloca i64, align 8
 // CHECK1-NEXT:    [[LC_CASTED:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[COMP_CASTED28:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS30:%.*]] = alloca [5 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS31:%.*]] = alloca [5 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS32:%.*]] = alloca [5 x i8*], align 8
+// CHECK1-NEXT:    [[COMP_CASTED31:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [5 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [5 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [5 x i8*], align 8
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32*, align 8
 // CHECK1-NEXT:    [[GBLC_CASTED:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[COMP_CASTED38:%.*]] = alloca i64, align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS40:%.*]] = alloca [2 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS41:%.*]] = alloca [2 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS42:%.*]] = alloca [2 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP43:%.*]] = alloca i32*, align 8
-// CHECK1-NEXT:    [[_TMP45:%.*]] = alloca i32*, align 8
+// CHECK1-NEXT:    [[COMP_CASTED42:%.*]] = alloca i64, align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS44:%.*]] = alloca [2 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS45:%.*]] = alloca [2 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS46:%.*]] = alloca [2 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP47:%.*]] = alloca i32*, align 8
+// CHECK1-NEXT:    [[_TMP49:%.*]] = alloca i32*, align 8
 // CHECK1-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
 // CHECK1-NEXT:    store i32 1, i32* [[COMP]], align 4
 // CHECK1-NEXT:    store i32 23, i32* [[LA]], align 4
@@ -357,223 +357,325 @@ void foo() {
 // CHECK1-NEXT:    store i8* null, i8** [[TMP6]], align 8
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l31.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l31.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l31(i64 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP11:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK1-NEXT:    [[TMP19:%.*]] = load i32, i32* [[COMP]], align 4
 // CHECK1-NEXT:    [[CONV2:%.*]] = bitcast i64* [[COMP_CASTED1]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP11]], i32* [[CONV2]], align 4
-// CHECK1-NEXT:    [[TMP12:%.*]] = load i64, i64* [[COMP_CASTED1]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP14:%.*]] = bitcast i8** [[TMP13]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP14]], align 8
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP12]], i64* [[TMP16]], align 8
-// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP17]], align 8
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l37.region_id, i32 1, i8** [[TMP18]], i8** [[TMP19]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
-// CHECK1-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK1:       omp_offload.failed6:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l37(i64 [[TMP12]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK1:       omp_offload.cont7:
-// CHECK1-NEXT:    [[TMP22:%.*]] = load i32, i32* [[LA]], align 4
-// CHECK1-NEXT:    [[CONV8:%.*]] = bitcast i64* [[LA_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP22]], i32* [[CONV8]], align 4
-// CHECK1-NEXT:    [[TMP23:%.*]] = load i64, i64* [[LA_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP24:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK1-NEXT:    [[CONV10:%.*]] = bitcast i64* [[COMP_CASTED9]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP24]], i32* [[CONV10]], align 4
-// CHECK1-NEXT:    [[TMP25:%.*]] = load i64, i64* [[COMP_CASTED9]], align 8
-// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP23]], i64* [[TMP27]], align 8
-// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP23]], i64* [[TMP29]], align 8
-// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP30]], align 8
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP32:%.*]] = bitcast i8** [[TMP31]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP25]], i64* [[TMP32]], align 8
-// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP34:%.*]] = bitcast i8** [[TMP33]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP25]], i64* [[TMP34]], align 8
-// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP35]], align 8
-// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 [[TMP19]], i32* [[CONV2]], align 4
+// CHECK1-NEXT:    [[TMP20:%.*]] = load i64, i64* [[COMP_CASTED1]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP20]], i64* [[TMP22]], align 8
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP20]], i64* [[TMP24]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[KERNEL_ARGS6:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP28]], align 4
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l37.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS6]])
+// CHECK1-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
+// CHECK1-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
+// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l37(i64 [[TMP20]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
+// CHECK1:       omp_offload.cont8:
 // CHECK1-NEXT:    [[TMP38:%.*]] = load i32, i32* [[LA]], align 4
-// CHECK1-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l46.region_id, i32 2, i8** [[TMP36]], i8** [[TMP37]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 [[TMP38]], i32 0)
-// CHECK1-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
-// CHECK1-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK1:       omp_offload.failed14:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l46(i64 [[TMP23]], i64 [[TMP25]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK1:       omp_offload.cont15:
-// CHECK1-NEXT:    [[TMP41:%.*]] = load i32, i32* [[LA]], align 4
-// CHECK1-NEXT:    [[CONV17:%.*]] = bitcast i64* [[LA_CASTED16]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP41]], i32* [[CONV17]], align 4
-// CHECK1-NEXT:    [[TMP42:%.*]] = load i64, i64* [[LA_CASTED16]], align 8
-// CHECK1-NEXT:    [[TMP43:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK1-NEXT:    [[CONV19:%.*]] = bitcast i64* [[COMP_CASTED18]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP43]], i32* [[CONV19]], align 4
-// CHECK1-NEXT:    [[TMP44:%.*]] = load i64, i64* [[COMP_CASTED18]], align 8
-// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP42]], i64* [[TMP46]], align 8
-// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
+// CHECK1-NEXT:    [[CONV9:%.*]] = bitcast i64* [[LA_CASTED]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP38]], i32* [[CONV9]], align 4
+// CHECK1-NEXT:    [[TMP39:%.*]] = load i64, i64* [[LA_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK1-NEXT:    [[CONV11:%.*]] = bitcast i64* [[COMP_CASTED10]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP40]], i32* [[CONV11]], align 4
+// CHECK1-NEXT:    [[TMP41:%.*]] = load i64, i64* [[COMP_CASTED10]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP39]], i64* [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP39]], i64* [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 1
 // CHECK1-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP42]], i64* [[TMP48]], align 8
-// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP49]], align 8
-// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP44]], i64* [[TMP51]], align 8
-// CHECK1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP44]], i64* [[TMP53]], align 8
-// CHECK1-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP54]], align 8
-// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP57:%.*]] = load i32, i32* [[LA]], align 4
-// CHECK1-NEXT:    [[TMP58:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l53.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP57]])
-// CHECK1-NEXT:    [[TMP59:%.*]] = icmp ne i32 [[TMP58]], 0
-// CHECK1-NEXT:    br i1 [[TMP59]], label [[OMP_OFFLOAD_FAILED23:%.*]], label [[OMP_OFFLOAD_CONT24:%.*]]
-// CHECK1:       omp_offload.failed23:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l53(i64 [[TMP42]], i64 [[TMP44]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT24]]
-// CHECK1:       omp_offload.cont24:
-// CHECK1-NEXT:    [[TMP60:%.*]] = load i32, i32* @Gbla, align 4
-// CHECK1-NEXT:    [[CONV25:%.*]] = bitcast i64* [[GBLA_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP60]], i32* [[CONV25]], align 4
-// CHECK1-NEXT:    [[TMP61:%.*]] = load i64, i64* [[GBLA_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP62:%.*]] = load i32, i32* [[A_ADDR]], align 4
-// CHECK1-NEXT:    [[CONV26:%.*]] = bitcast i64* [[A_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP62]], i32* [[CONV26]], align 4
-// CHECK1-NEXT:    [[TMP63:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP64:%.*]] = load i64, i64* @Gblb, align 8
-// CHECK1-NEXT:    store i64 [[TMP64]], i64* [[GBLB_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP65:%.*]] = load i64, i64* [[GBLB_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP66:%.*]] = load float, float* [[LC]], align 4
-// CHECK1-NEXT:    [[CONV27:%.*]] = bitcast i64* [[LC_CASTED]] to float*
-// CHECK1-NEXT:    store float [[TMP66]], float* [[CONV27]], align 4
-// CHECK1-NEXT:    [[TMP67:%.*]] = load i64, i64* [[LC_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP68:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK1-NEXT:    [[CONV29:%.*]] = bitcast i64* [[COMP_CASTED28]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP68]], i32* [[CONV29]], align 4
-// CHECK1-NEXT:    [[TMP69:%.*]] = load i64, i64* [[COMP_CASTED28]], align 8
-// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP61]], i64* [[TMP71]], align 8
-// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP61]], i64* [[TMP73]], align 8
-// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP74]], align 8
-// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP63]], i64* [[TMP76]], align 8
-// CHECK1-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP63]], i64* [[TMP78]], align 8
-// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP79]], align 8
-// CHECK1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP65]], i64* [[TMP81]], align 8
-// CHECK1-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP65]], i64* [[TMP83]], align 8
-// CHECK1-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP84]], align 8
-// CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP67]], i64* [[TMP86]], align 8
-// CHECK1-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 3
-// CHECK1-NEXT:    [[TMP88:%.*]] = bitcast i8** [[TMP87]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP67]], i64* [[TMP88]], align 8
-// CHECK1-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 3
-// CHECK1-NEXT:    store i8* null, i8** [[TMP89]], align 8
-// CHECK1-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP69]], i64* [[TMP91]], align 8
-// CHECK1-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 4
-// CHECK1-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP69]], i64* [[TMP93]], align 8
-// CHECK1-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS32]], i64 0, i64 4
-// CHECK1-NEXT:    store i8* null, i8** [[TMP94]], align 8
-// CHECK1-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS30]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS31]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP97:%.*]] = load i32, i32* @Gbla, align 4
-// CHECK1-NEXT:    [[TMP98:%.*]] = load i32, i32* [[A_ADDR]], align 4
-// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP97]], [[TMP98]]
-// CHECK1-NEXT:    [[TMP99:%.*]] = load i64, i64* @Gblb, align 8
-// CHECK1-NEXT:    [[TMP100:%.*]] = load float, float* [[LC]], align 4
-// CHECK1-NEXT:    [[CONV33:%.*]] = fptosi float [[TMP100]] to i64
-// CHECK1-NEXT:    [[ADD34:%.*]] = add nsw i64 [[TMP99]], [[CONV33]]
-// CHECK1-NEXT:    [[TMP101:%.*]] = trunc i64 [[ADD34]] to i32
-// CHECK1-NEXT:    [[TMP102:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l62.region_id, i32 5, i8** [[TMP95]], i8** [[TMP96]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 [[ADD]], i32 [[TMP101]])
-// CHECK1-NEXT:    [[TMP103:%.*]] = icmp ne i32 [[TMP102]], 0
-// CHECK1-NEXT:    br i1 [[TMP103]], label [[OMP_OFFLOAD_FAILED35:%.*]], label [[OMP_OFFLOAD_CONT36:%.*]]
-// CHECK1:       omp_offload.failed35:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l62(i64 [[TMP61]], i64 [[TMP63]], i64 [[TMP65]], i64 [[TMP67]], i64 [[TMP69]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT36]]
-// CHECK1:       omp_offload.cont36:
-// CHECK1-NEXT:    [[TMP104:%.*]] = load i32*, i32** @Gblc, align 8
-// CHECK1-NEXT:    store i32* [[TMP104]], i32** [[TMP]], align 8
-// CHECK1-NEXT:    [[TMP105:%.*]] = load i32, i32* @Gbla, align 4
-// CHECK1-NEXT:    [[CONV37:%.*]] = bitcast i64* [[GBLC_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP105]], i32* [[CONV37]], align 4
-// CHECK1-NEXT:    [[TMP106:%.*]] = load i64, i64* [[GBLC_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP107:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK1-NEXT:    [[CONV39:%.*]] = bitcast i64* [[COMP_CASTED38]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP107]], i32* [[CONV39]], align 4
-// CHECK1-NEXT:    [[TMP108:%.*]] = load i64, i64* [[COMP_CASTED38]], align 8
-// CHECK1-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS40]], i32 0, i32 0
+// CHECK1-NEXT:    store i64 [[TMP41]], i64* [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP41]], i64* [[TMP50]], align 8
+// CHECK1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS14]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP51]], align 8
+// CHECK1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS13]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP54:%.*]] = load i32, i32* [[LA]], align 4
+// CHECK1-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP55]], align 4
+// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP56]], align 4
+// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP52]], i8*** [[TMP57]], align 8
+// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP53]], i8*** [[TMP58]], align 8
+// CHECK1-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP59]], align 8
+// CHECK1-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP60]], align 8
+// CHECK1-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP61]], align 8
+// CHECK1-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP62]], align 8
+// CHECK1-NEXT:    [[TMP63:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP54]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l46.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK1-NEXT:    [[TMP64:%.*]] = icmp ne i32 [[TMP63]], 0
+// CHECK1-NEXT:    br i1 [[TMP64]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK1:       omp_offload.failed16:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l46(i64 [[TMP39]], i64 [[TMP41]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK1:       omp_offload.cont17:
+// CHECK1-NEXT:    [[TMP65:%.*]] = load i32, i32* [[LA]], align 4
+// CHECK1-NEXT:    [[CONV19:%.*]] = bitcast i64* [[LA_CASTED18]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP65]], i32* [[CONV19]], align 4
+// CHECK1-NEXT:    [[TMP66:%.*]] = load i64, i64* [[LA_CASTED18]], align 8
+// CHECK1-NEXT:    [[TMP67:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK1-NEXT:    [[CONV21:%.*]] = bitcast i64* [[COMP_CASTED20]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP67]], i32* [[CONV21]], align 4
+// CHECK1-NEXT:    [[TMP68:%.*]] = load i64, i64* [[COMP_CASTED20]], align 8
+// CHECK1-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP66]], i64* [[TMP70]], align 8
+// CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP66]], i64* [[TMP72]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP73]], align 8
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP68]], i64* [[TMP75]], align 8
+// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP68]], i64* [[TMP77]], align 8
+// CHECK1-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP78]], align 8
+// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP81:%.*]] = load i32, i32* [[LA]], align 4
+// CHECK1-NEXT:    [[KERNEL_ARGS25:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS25]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP82]], align 4
+// CHECK1-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS25]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP83]], align 4
+// CHECK1-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS25]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP79]], i8*** [[TMP84]], align 8
+// CHECK1-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS25]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP80]], i8*** [[TMP85]], align 8
+// CHECK1-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS25]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP86]], align 8
+// CHECK1-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS25]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP87]], align 8
+// CHECK1-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS25]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP88]], align 8
+// CHECK1-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS25]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP89]], align 8
+// CHECK1-NEXT:    [[TMP90:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 [[TMP81]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l53.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS25]])
+// CHECK1-NEXT:    [[TMP91:%.*]] = icmp ne i32 [[TMP90]], 0
+// CHECK1-NEXT:    br i1 [[TMP91]], label [[OMP_OFFLOAD_FAILED26:%.*]], label [[OMP_OFFLOAD_CONT27:%.*]]
+// CHECK1:       omp_offload.failed26:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l53(i64 [[TMP66]], i64 [[TMP68]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT27]]
+// CHECK1:       omp_offload.cont27:
+// CHECK1-NEXT:    [[TMP92:%.*]] = load i32, i32* @Gbla, align 4
+// CHECK1-NEXT:    [[CONV28:%.*]] = bitcast i64* [[GBLA_CASTED]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP92]], i32* [[CONV28]], align 4
+// CHECK1-NEXT:    [[TMP93:%.*]] = load i64, i64* [[GBLA_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP94:%.*]] = load i32, i32* [[A_ADDR]], align 4
+// CHECK1-NEXT:    [[CONV29:%.*]] = bitcast i64* [[A_CASTED]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP94]], i32* [[CONV29]], align 4
+// CHECK1-NEXT:    [[TMP95:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP96:%.*]] = load i64, i64* @Gblb, align 8
+// CHECK1-NEXT:    store i64 [[TMP96]], i64* [[GBLB_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP97:%.*]] = load i64, i64* [[GBLB_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP98:%.*]] = load float, float* [[LC]], align 4
+// CHECK1-NEXT:    [[CONV30:%.*]] = bitcast i64* [[LC_CASTED]] to float*
+// CHECK1-NEXT:    store float [[TMP98]], float* [[CONV30]], align 4
+// CHECK1-NEXT:    [[TMP99:%.*]] = load i64, i64* [[LC_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP100:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK1-NEXT:    [[CONV32:%.*]] = bitcast i64* [[COMP_CASTED31]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP100]], i32* [[CONV32]], align 4
+// CHECK1-NEXT:    [[TMP101:%.*]] = load i64, i64* [[COMP_CASTED31]], align 8
+// CHECK1-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP93]], i64* [[TMP103]], align 8
+// CHECK1-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP93]], i64* [[TMP105]], align 8
+// CHECK1-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP106]], align 8
+// CHECK1-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP95]], i64* [[TMP108]], align 8
+// CHECK1-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 1
 // CHECK1-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP106]], i64* [[TMP110]], align 8
-// CHECK1-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS41]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP106]], i64* [[TMP112]], align 8
-// CHECK1-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS42]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP113]], align 8
-// CHECK1-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS40]], i32 0, i32 1
+// CHECK1-NEXT:    store i64 [[TMP95]], i64* [[TMP110]], align 8
+// CHECK1-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP111]], align 8
+// CHECK1-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP97]], i64* [[TMP113]], align 8
+// CHECK1-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 2
 // CHECK1-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP108]], i64* [[TMP115]], align 8
-// CHECK1-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS41]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP108]], i64* [[TMP117]], align 8
-// CHECK1-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS42]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP118]], align 8
-// CHECK1-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS40]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS41]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP121:%.*]] = load i32*, i32** @Gblc, align 8
-// CHECK1-NEXT:    store i32* [[TMP121]], i32** [[_TMP43]], align 8
-// CHECK1-NEXT:    [[TMP122:%.*]] = load i32, i32* @Gbla, align 4
-// CHECK1-NEXT:    [[ADD44:%.*]] = add nsw i32 [[TMP122]], 1
-// CHECK1-NEXT:    [[TMP123:%.*]] = load i32*, i32** @Gblc, align 8
-// CHECK1-NEXT:    store i32* [[TMP123]], i32** [[_TMP45]], align 8
-// CHECK1-NEXT:    [[TMP124:%.*]] = load i32, i32* @Gbla, align 4
-// CHECK1-NEXT:    [[ADD46:%.*]] = add nsw i32 [[TMP124]], 2
-// CHECK1-NEXT:    [[TMP125:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l71.region_id, i32 2, i8** [[TMP119]], i8** [[TMP120]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 [[ADD44]], i32 [[ADD46]])
-// CHECK1-NEXT:    [[TMP126:%.*]] = icmp ne i32 [[TMP125]], 0
-// CHECK1-NEXT:    br i1 [[TMP126]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
-// CHECK1:       omp_offload.failed47:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l71(i64 [[TMP106]], i64 [[TMP108]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
-// CHECK1:       omp_offload.cont48:
-// CHECK1-NEXT:    [[TMP127:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP127]]
+// CHECK1-NEXT:    store i64 [[TMP97]], i64* [[TMP115]], align 8
+// CHECK1-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i64 0, i64 2
+// CHECK1-NEXT:    store i8* null, i8** [[TMP116]], align 8
+// CHECK1-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP99]], i64* [[TMP118]], align 8
+// CHECK1-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 3
+// CHECK1-NEXT:    [[TMP120:%.*]] = bitcast i8** [[TMP119]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP99]], i64* [[TMP120]], align 8
+// CHECK1-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i64 0, i64 3
+// CHECK1-NEXT:    store i8* null, i8** [[TMP121]], align 8
+// CHECK1-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP123:%.*]] = bitcast i8** [[TMP122]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP101]], i64* [[TMP123]], align 8
+// CHECK1-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 4
+// CHECK1-NEXT:    [[TMP125:%.*]] = bitcast i8** [[TMP124]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP101]], i64* [[TMP125]], align 8
+// CHECK1-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i64 0, i64 4
+// CHECK1-NEXT:    store i8* null, i8** [[TMP126]], align 8
+// CHECK1-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP129:%.*]] = load i32, i32* @Gbla, align 4
+// CHECK1-NEXT:    [[TMP130:%.*]] = load i32, i32* [[A_ADDR]], align 4
+// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP129]], [[TMP130]]
+// CHECK1-NEXT:    [[TMP131:%.*]] = load i64, i64* @Gblb, align 8
+// CHECK1-NEXT:    [[TMP132:%.*]] = load float, float* [[LC]], align 4
+// CHECK1-NEXT:    [[CONV36:%.*]] = fptosi float [[TMP132]] to i64
+// CHECK1-NEXT:    [[ADD37:%.*]] = add nsw i64 [[TMP131]], [[CONV36]]
+// CHECK1-NEXT:    [[TMP133:%.*]] = trunc i64 [[ADD37]] to i32
+// CHECK1-NEXT:    [[KERNEL_ARGS38:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP134]], align 4
+// CHECK1-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP135]], align 4
+// CHECK1-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP127]], i8*** [[TMP136]], align 8
+// CHECK1-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP128]], i8*** [[TMP137]], align 8
+// CHECK1-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP138]], align 8
+// CHECK1-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP139]], align 8
+// CHECK1-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP140]], align 8
+// CHECK1-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP141]], align 8
+// CHECK1-NEXT:    [[TMP142:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[ADD]], i32 [[TMP133]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l62.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS38]])
+// CHECK1-NEXT:    [[TMP143:%.*]] = icmp ne i32 [[TMP142]], 0
+// CHECK1-NEXT:    br i1 [[TMP143]], label [[OMP_OFFLOAD_FAILED39:%.*]], label [[OMP_OFFLOAD_CONT40:%.*]]
+// CHECK1:       omp_offload.failed39:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l62(i64 [[TMP93]], i64 [[TMP95]], i64 [[TMP97]], i64 [[TMP99]], i64 [[TMP101]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT40]]
+// CHECK1:       omp_offload.cont40:
+// CHECK1-NEXT:    [[TMP144:%.*]] = load i32*, i32** @Gblc, align 8
+// CHECK1-NEXT:    store i32* [[TMP144]], i32** [[TMP]], align 8
+// CHECK1-NEXT:    [[TMP145:%.*]] = load i32, i32* @Gbla, align 4
+// CHECK1-NEXT:    [[CONV41:%.*]] = bitcast i64* [[GBLC_CASTED]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP145]], i32* [[CONV41]], align 4
+// CHECK1-NEXT:    [[TMP146:%.*]] = load i64, i64* [[GBLC_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP147:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK1-NEXT:    [[CONV43:%.*]] = bitcast i64* [[COMP_CASTED42]] to i32*
+// CHECK1-NEXT:    store i32 [[TMP147]], i32* [[CONV43]], align 4
+// CHECK1-NEXT:    [[TMP148:%.*]] = load i64, i64* [[COMP_CASTED42]], align 8
+// CHECK1-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP150:%.*]] = bitcast i8** [[TMP149]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP146]], i64* [[TMP150]], align 8
+// CHECK1-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP146]], i64* [[TMP152]], align 8
+// CHECK1-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP153]], align 8
+// CHECK1-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP155:%.*]] = bitcast i8** [[TMP154]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP148]], i64* [[TMP155]], align 8
+// CHECK1-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP157:%.*]] = bitcast i8** [[TMP156]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP148]], i64* [[TMP157]], align 8
+// CHECK1-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS46]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP158]], align 8
+// CHECK1-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS44]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP160:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS45]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP161:%.*]] = load i32*, i32** @Gblc, align 8
+// CHECK1-NEXT:    store i32* [[TMP161]], i32** [[_TMP47]], align 8
+// CHECK1-NEXT:    [[TMP162:%.*]] = load i32, i32* @Gbla, align 4
+// CHECK1-NEXT:    [[ADD48:%.*]] = add nsw i32 [[TMP162]], 1
+// CHECK1-NEXT:    [[TMP163:%.*]] = load i32*, i32** @Gblc, align 8
+// CHECK1-NEXT:    store i32* [[TMP163]], i32** [[_TMP49]], align 8
+// CHECK1-NEXT:    [[TMP164:%.*]] = load i32, i32* @Gbla, align 4
+// CHECK1-NEXT:    [[ADD50:%.*]] = add nsw i32 [[TMP164]], 2
+// CHECK1-NEXT:    [[KERNEL_ARGS51:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS51]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP165]], align 4
+// CHECK1-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS51]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP166]], align 4
+// CHECK1-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS51]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP159]], i8*** [[TMP167]], align 8
+// CHECK1-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS51]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP160]], i8*** [[TMP168]], align 8
+// CHECK1-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS51]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP169]], align 8
+// CHECK1-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS51]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP170]], align 8
+// CHECK1-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS51]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP171]], align 8
+// CHECK1-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS51]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP172]], align 8
+// CHECK1-NEXT:    [[TMP173:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[ADD48]], i32 [[ADD50]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l71.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS51]])
+// CHECK1-NEXT:    [[TMP174:%.*]] = icmp ne i32 [[TMP173]], 0
+// CHECK1-NEXT:    br i1 [[TMP174]], label [[OMP_OFFLOAD_FAILED52:%.*]], label [[OMP_OFFLOAD_CONT53:%.*]]
+// CHECK1:       omp_offload.failed52:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l71(i64 [[TMP146]], i64 [[TMP148]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT53]]
+// CHECK1:       omp_offload.cont53:
+// CHECK1-NEXT:    [[TMP175:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP175]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l31
@@ -806,30 +908,30 @@ void foo() {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS3:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[LA_CASTED:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[COMP_CASTED7:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 4
-// CHECK3-NEXT:    [[LA_CASTED13:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[COMP_CASTED14:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [2 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [2 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [2 x i8*], align 4
+// CHECK3-NEXT:    [[COMP_CASTED8:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS9:%.*]] = alloca [2 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS10:%.*]] = alloca [2 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS11:%.*]] = alloca [2 x i8*], align 4
+// CHECK3-NEXT:    [[LA_CASTED15:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[COMP_CASTED16:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [2 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [2 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [2 x i8*], align 4
 // CHECK3-NEXT:    [[GBLA_CASTED:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[LC_CASTED:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[COMP_CASTED20:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [5 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [5 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [5 x i8*], align 4
+// CHECK3-NEXT:    [[COMP_CASTED23:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [5 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [5 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [5 x i8*], align 4
 // CHECK3-NEXT:    [[TMP:%.*]] = alloca i32*, align 4
 // CHECK3-NEXT:    [[GBLC_CASTED:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[COMP_CASTED28:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS29:%.*]] = alloca [2 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS30:%.*]] = alloca [2 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS31:%.*]] = alloca [2 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP32:%.*]] = alloca i32*, align 4
-// CHECK3-NEXT:    [[_TMP34:%.*]] = alloca i32*, align 4
+// CHECK3-NEXT:    [[COMP_CASTED32:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS33:%.*]] = alloca [2 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS34:%.*]] = alloca [2 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS35:%.*]] = alloca [2 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP36:%.*]] = alloca i32*, align 4
+// CHECK3-NEXT:    [[_TMP38:%.*]] = alloca i32*, align 4
 // CHECK3-NEXT:    store i32 [[A]], i32* [[A_ADDR]], align 4
 // CHECK3-NEXT:    store i32 1, i32* [[COMP]], align 4
 // CHECK3-NEXT:    store i32 23, i32* [[LA]], align 4
@@ -847,210 +949,312 @@ void foo() {
 // CHECK3-NEXT:    store i8* null, i8** [[TMP6]], align 4
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l31.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l31.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK3-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l31(i32 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP11:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK3-NEXT:    store i32 [[TMP11]], i32* [[COMP_CASTED1]], align 4
-// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[COMP_CASTED1]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP14:%.*]] = bitcast i8** [[TMP13]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP14]], align 4
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP12]], i32* [[TMP16]], align 4
-// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP17]], align 4
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l37.region_id, i32 1, i8** [[TMP18]], i8** [[TMP19]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
-// CHECK3-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK3:       omp_offload.failed5:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l37(i32 [[TMP12]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK3:       omp_offload.cont6:
-// CHECK3-NEXT:    [[TMP22:%.*]] = load i32, i32* [[LA]], align 4
-// CHECK3-NEXT:    store i32 [[TMP22]], i32* [[LA_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = load i32, i32* [[LA_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP24:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK3-NEXT:    store i32 [[TMP24]], i32* [[COMP_CASTED7]], align 4
-// CHECK3-NEXT:    [[TMP25:%.*]] = load i32, i32* [[COMP_CASTED7]], align 4
-// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP23]], i32* [[TMP27]], align 4
-// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP23]], i32* [[TMP29]], align 4
-// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP30]], align 4
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP32:%.*]] = bitcast i8** [[TMP31]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP25]], i32* [[TMP32]], align 4
-// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP34:%.*]] = bitcast i8** [[TMP33]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP25]], i32* [[TMP34]], align 4
-// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP35]], align 4
-// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP19:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK3-NEXT:    store i32 [[TMP19]], i32* [[COMP_CASTED1]], align 4
+// CHECK3-NEXT:    [[TMP20:%.*]] = load i32, i32* [[COMP_CASTED1]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP20]], i32* [[TMP22]], align 4
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP24:%.*]] = bitcast i8** [[TMP23]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP20]], i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP26]], i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l37.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK3-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
+// CHECK3-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK3:       omp_offload.failed6:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l37(i32 [[TMP20]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK3:       omp_offload.cont7:
 // CHECK3-NEXT:    [[TMP38:%.*]] = load i32, i32* [[LA]], align 4
-// CHECK3-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l46.region_id, i32 2, i8** [[TMP36]], i8** [[TMP37]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 [[TMP38]], i32 0)
-// CHECK3-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
-// CHECK3-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK3:       omp_offload.failed11:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l46(i32 [[TMP23]], i32 [[TMP25]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK3:       omp_offload.cont12:
-// CHECK3-NEXT:    [[TMP41:%.*]] = load i32, i32* [[LA]], align 4
-// CHECK3-NEXT:    store i32 [[TMP41]], i32* [[LA_CASTED13]], align 4
-// CHECK3-NEXT:    [[TMP42:%.*]] = load i32, i32* [[LA_CASTED13]], align 4
-// CHECK3-NEXT:    [[TMP43:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK3-NEXT:    store i32 [[TMP43]], i32* [[COMP_CASTED14]], align 4
-// CHECK3-NEXT:    [[TMP44:%.*]] = load i32, i32* [[COMP_CASTED14]], align 4
-// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP42]], i32* [[TMP46]], align 4
-// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 [[TMP38]], i32* [[LA_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = load i32, i32* [[LA_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP40:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK3-NEXT:    store i32 [[TMP40]], i32* [[COMP_CASTED8]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = load i32, i32* [[COMP_CASTED8]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP39]], i32* [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP39]], i32* [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 1
 // CHECK3-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP42]], i32* [[TMP48]], align 4
-// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP49]], align 4
-// CHECK3-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP44]], i32* [[TMP51]], align 4
-// CHECK3-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP44]], i32* [[TMP53]], align 4
-// CHECK3-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP54]], align 4
-// CHECK3-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP57:%.*]] = load i32, i32* [[LA]], align 4
-// CHECK3-NEXT:    [[TMP58:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l53.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP57]])
-// CHECK3-NEXT:    [[TMP59:%.*]] = icmp ne i32 [[TMP58]], 0
-// CHECK3-NEXT:    br i1 [[TMP59]], label [[OMP_OFFLOAD_FAILED18:%.*]], label [[OMP_OFFLOAD_CONT19:%.*]]
-// CHECK3:       omp_offload.failed18:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l53(i32 [[TMP42]], i32 [[TMP44]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT19]]
-// CHECK3:       omp_offload.cont19:
-// CHECK3-NEXT:    [[TMP60:%.*]] = load i32, i32* @Gbla, align 4
-// CHECK3-NEXT:    store i32 [[TMP60]], i32* [[GBLA_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP61:%.*]] = load i32, i32* [[GBLA_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP62:%.*]] = load i32, i32* [[A_ADDR]], align 4
-// CHECK3-NEXT:    store i32 [[TMP62]], i32* [[A_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP63:%.*]] = load i32, i32* [[A_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP64:%.*]] = load float, float* [[LC]], align 4
+// CHECK3-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP41]], i32* [[TMP50]], align 4
+// CHECK3-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP51]], align 4
+// CHECK3-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP54:%.*]] = load i32, i32* [[LA]], align 4
+// CHECK3-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP55]], align 4
+// CHECK3-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP56]], align 4
+// CHECK3-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP52]], i8*** [[TMP57]], align 4
+// CHECK3-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP53]], i8*** [[TMP58]], align 4
+// CHECK3-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP59]], align 4
+// CHECK3-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP60]], align 4
+// CHECK3-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP61]], align 4
+// CHECK3-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP62]], align 4
+// CHECK3-NEXT:    [[TMP63:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP54]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l46.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK3-NEXT:    [[TMP64:%.*]] = icmp ne i32 [[TMP63]], 0
+// CHECK3-NEXT:    br i1 [[TMP64]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK3:       omp_offload.failed13:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l46(i32 [[TMP39]], i32 [[TMP41]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK3:       omp_offload.cont14:
+// CHECK3-NEXT:    [[TMP65:%.*]] = load i32, i32* [[LA]], align 4
+// CHECK3-NEXT:    store i32 [[TMP65]], i32* [[LA_CASTED15]], align 4
+// CHECK3-NEXT:    [[TMP66:%.*]] = load i32, i32* [[LA_CASTED15]], align 4
+// CHECK3-NEXT:    [[TMP67:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK3-NEXT:    store i32 [[TMP67]], i32* [[COMP_CASTED16]], align 4
+// CHECK3-NEXT:    [[TMP68:%.*]] = load i32, i32* [[COMP_CASTED16]], align 4
+// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP66]], i32* [[TMP70]], align 4
+// CHECK3-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP66]], i32* [[TMP72]], align 4
+// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP73]], align 4
+// CHECK3-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP68]], i32* [[TMP75]], align 4
+// CHECK3-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP68]], i32* [[TMP77]], align 4
+// CHECK3-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP78]], align 4
+// CHECK3-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP81:%.*]] = load i32, i32* [[LA]], align 4
+// CHECK3-NEXT:    [[KERNEL_ARGS20:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP82]], align 4
+// CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP83]], align 4
+// CHECK3-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP79]], i8*** [[TMP84]], align 4
+// CHECK3-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP80]], i8*** [[TMP85]], align 4
+// CHECK3-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP86]], align 4
+// CHECK3-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP87]], align 4
+// CHECK3-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP88]], align 4
+// CHECK3-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP89]], align 4
+// CHECK3-NEXT:    [[TMP90:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 [[TMP81]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l53.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS20]])
+// CHECK3-NEXT:    [[TMP91:%.*]] = icmp ne i32 [[TMP90]], 0
+// CHECK3-NEXT:    br i1 [[TMP91]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
+// CHECK3:       omp_offload.failed21:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l53(i32 [[TMP66]], i32 [[TMP68]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
+// CHECK3:       omp_offload.cont22:
+// CHECK3-NEXT:    [[TMP92:%.*]] = load i32, i32* @Gbla, align 4
+// CHECK3-NEXT:    store i32 [[TMP92]], i32* [[GBLA_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP93:%.*]] = load i32, i32* [[GBLA_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP94:%.*]] = load i32, i32* [[A_ADDR]], align 4
+// CHECK3-NEXT:    store i32 [[TMP94]], i32* [[A_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP95:%.*]] = load i32, i32* [[A_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP96:%.*]] = load float, float* [[LC]], align 4
 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i32* [[LC_CASTED]] to float*
-// CHECK3-NEXT:    store float [[TMP64]], float* [[CONV]], align 4
-// CHECK3-NEXT:    [[TMP65:%.*]] = load i32, i32* [[LC_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP66:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK3-NEXT:    store i32 [[TMP66]], i32* [[COMP_CASTED20]], align 4
-// CHECK3-NEXT:    [[TMP67:%.*]] = load i32, i32* [[COMP_CASTED20]], align 4
-// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP61]], i32* [[TMP69]], align 4
-// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP61]], i32* [[TMP71]], align 4
-// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP72]], align 4
-// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP74:%.*]] = bitcast i8** [[TMP73]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP63]], i32* [[TMP74]], align 4
-// CHECK3-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP63]], i32* [[TMP76]], align 4
-// CHECK3-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP77]], align 4
-// CHECK3-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP79:%.*]] = bitcast i8** [[TMP78]] to i64**
-// CHECK3-NEXT:    store i64* @Gblb, i64** [[TMP79]], align 4
-// CHECK3-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP81:%.*]] = bitcast i8** [[TMP80]] to i64**
-// CHECK3-NEXT:    store i64* @Gblb, i64** [[TMP81]], align 4
-// CHECK3-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP82]], align 4
-// CHECK3-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP65]], i32* [[TMP84]], align 4
-// CHECK3-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
-// CHECK3-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP65]], i32* [[TMP86]], align 4
-// CHECK3-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 3
-// CHECK3-NEXT:    store i8* null, i8** [[TMP87]], align 4
-// CHECK3-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP67]], i32* [[TMP89]], align 4
-// CHECK3-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 4
-// CHECK3-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP67]], i32* [[TMP91]], align 4
-// CHECK3-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i32 0, i32 4
-// CHECK3-NEXT:    store i8* null, i8** [[TMP92]], align 4
-// CHECK3-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP95:%.*]] = load i32, i32* @Gbla, align 4
-// CHECK3-NEXT:    [[TMP96:%.*]] = load i32, i32* [[A_ADDR]], align 4
-// CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP95]], [[TMP96]]
-// CHECK3-NEXT:    [[TMP97:%.*]] = load i64, i64* @Gblb, align 8
-// CHECK3-NEXT:    [[TMP98:%.*]] = load float, float* [[LC]], align 4
-// CHECK3-NEXT:    [[CONV24:%.*]] = fptosi float [[TMP98]] to i64
-// CHECK3-NEXT:    [[ADD25:%.*]] = add nsw i64 [[TMP97]], [[CONV24]]
-// CHECK3-NEXT:    [[TMP99:%.*]] = trunc i64 [[ADD25]] to i32
-// CHECK3-NEXT:    [[TMP100:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l62.region_id, i32 5, i8** [[TMP93]], i8** [[TMP94]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 [[ADD]], i32 [[TMP99]])
-// CHECK3-NEXT:    [[TMP101:%.*]] = icmp ne i32 [[TMP100]], 0
-// CHECK3-NEXT:    br i1 [[TMP101]], label [[OMP_OFFLOAD_FAILED26:%.*]], label [[OMP_OFFLOAD_CONT27:%.*]]
-// CHECK3:       omp_offload.failed26:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l62(i32 [[TMP61]], i32 [[TMP63]], i64* @Gblb, i32 [[TMP65]], i32 [[TMP67]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT27]]
-// CHECK3:       omp_offload.cont27:
-// CHECK3-NEXT:    [[TMP102:%.*]] = load i32*, i32** @Gblc, align 4
-// CHECK3-NEXT:    store i32* [[TMP102]], i32** [[TMP]], align 4
-// CHECK3-NEXT:    [[TMP103:%.*]] = load i32, i32* @Gbla, align 4
-// CHECK3-NEXT:    store i32 [[TMP103]], i32* [[GBLC_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP104:%.*]] = load i32, i32* [[GBLC_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP105:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK3-NEXT:    store i32 [[TMP105]], i32* [[COMP_CASTED28]], align 4
-// CHECK3-NEXT:    [[TMP106:%.*]] = load i32, i32* [[COMP_CASTED28]], align 4
-// CHECK3-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 0
+// CHECK3-NEXT:    store float [[TMP96]], float* [[CONV]], align 4
+// CHECK3-NEXT:    [[TMP97:%.*]] = load i32, i32* [[LC_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP98:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK3-NEXT:    store i32 [[TMP98]], i32* [[COMP_CASTED23]], align 4
+// CHECK3-NEXT:    [[TMP99:%.*]] = load i32, i32* [[COMP_CASTED23]], align 4
+// CHECK3-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP93]], i32* [[TMP101]], align 4
+// CHECK3-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP93]], i32* [[TMP103]], align 4
+// CHECK3-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP104]], align 4
+// CHECK3-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP106:%.*]] = bitcast i8** [[TMP105]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP95]], i32* [[TMP106]], align 4
+// CHECK3-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 1
 // CHECK3-NEXT:    [[TMP108:%.*]] = bitcast i8** [[TMP107]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP104]], i32* [[TMP108]], align 4
-// CHECK3-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP104]], i32* [[TMP110]], align 4
-// CHECK3-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP111]], align 4
-// CHECK3-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP106]], i32* [[TMP113]], align 4
-// CHECK3-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP106]], i32* [[TMP115]], align 4
-// CHECK3-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS31]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP116]], align 4
-// CHECK3-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS29]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS30]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP119:%.*]] = load i32*, i32** @Gblc, align 4
-// CHECK3-NEXT:    store i32* [[TMP119]], i32** [[_TMP32]], align 4
-// CHECK3-NEXT:    [[TMP120:%.*]] = load i32, i32* @Gbla, align 4
-// CHECK3-NEXT:    [[ADD33:%.*]] = add nsw i32 [[TMP120]], 1
-// CHECK3-NEXT:    [[TMP121:%.*]] = load i32*, i32** @Gblc, align 4
-// CHECK3-NEXT:    store i32* [[TMP121]], i32** [[_TMP34]], align 4
-// CHECK3-NEXT:    [[TMP122:%.*]] = load i32, i32* @Gbla, align 4
-// CHECK3-NEXT:    [[ADD35:%.*]] = add nsw i32 [[TMP122]], 2
-// CHECK3-NEXT:    [[TMP123:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l71.region_id, i32 2, i8** [[TMP117]], i8** [[TMP118]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 [[ADD33]], i32 [[ADD35]])
-// CHECK3-NEXT:    [[TMP124:%.*]] = icmp ne i32 [[TMP123]], 0
-// CHECK3-NEXT:    br i1 [[TMP124]], label [[OMP_OFFLOAD_FAILED36:%.*]], label [[OMP_OFFLOAD_CONT37:%.*]]
-// CHECK3:       omp_offload.failed36:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l71(i32 [[TMP104]], i32 [[TMP106]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT37]]
-// CHECK3:       omp_offload.cont37:
-// CHECK3-NEXT:    [[TMP125:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP125]]
+// CHECK3-NEXT:    store i32 [[TMP95]], i32* [[TMP108]], align 4
+// CHECK3-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP109]], align 4
+// CHECK3-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP111:%.*]] = bitcast i8** [[TMP110]] to i64**
+// CHECK3-NEXT:    store i64* @Gblb, i64** [[TMP111]], align 4
+// CHECK3-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP113:%.*]] = bitcast i8** [[TMP112]] to i64**
+// CHECK3-NEXT:    store i64* @Gblb, i64** [[TMP113]], align 4
+// CHECK3-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 2
+// CHECK3-NEXT:    store i8* null, i8** [[TMP114]], align 4
+// CHECK3-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP116:%.*]] = bitcast i8** [[TMP115]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP97]], i32* [[TMP116]], align 4
+// CHECK3-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 3
+// CHECK3-NEXT:    [[TMP118:%.*]] = bitcast i8** [[TMP117]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP97]], i32* [[TMP118]], align 4
+// CHECK3-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 3
+// CHECK3-NEXT:    store i8* null, i8** [[TMP119]], align 4
+// CHECK3-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP121:%.*]] = bitcast i8** [[TMP120]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP99]], i32* [[TMP121]], align 4
+// CHECK3-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 4
+// CHECK3-NEXT:    [[TMP123:%.*]] = bitcast i8** [[TMP122]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP99]], i32* [[TMP123]], align 4
+// CHECK3-NEXT:    [[TMP124:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 4
+// CHECK3-NEXT:    store i8* null, i8** [[TMP124]], align 4
+// CHECK3-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP126:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP127:%.*]] = load i32, i32* @Gbla, align 4
+// CHECK3-NEXT:    [[TMP128:%.*]] = load i32, i32* [[A_ADDR]], align 4
+// CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP127]], [[TMP128]]
+// CHECK3-NEXT:    [[TMP129:%.*]] = load i64, i64* @Gblb, align 8
+// CHECK3-NEXT:    [[TMP130:%.*]] = load float, float* [[LC]], align 4
+// CHECK3-NEXT:    [[CONV27:%.*]] = fptosi float [[TMP130]] to i64
+// CHECK3-NEXT:    [[ADD28:%.*]] = add nsw i64 [[TMP129]], [[CONV27]]
+// CHECK3-NEXT:    [[TMP131:%.*]] = trunc i64 [[ADD28]] to i32
+// CHECK3-NEXT:    [[KERNEL_ARGS29:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP132]], align 4
+// CHECK3-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP133]], align 4
+// CHECK3-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP125]], i8*** [[TMP134]], align 4
+// CHECK3-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP126]], i8*** [[TMP135]], align 4
+// CHECK3-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP136]], align 4
+// CHECK3-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP137]], align 4
+// CHECK3-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP138]], align 4
+// CHECK3-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP139]], align 4
+// CHECK3-NEXT:    [[TMP140:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[ADD]], i32 [[TMP131]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l62.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS29]])
+// CHECK3-NEXT:    [[TMP141:%.*]] = icmp ne i32 [[TMP140]], 0
+// CHECK3-NEXT:    br i1 [[TMP141]], label [[OMP_OFFLOAD_FAILED30:%.*]], label [[OMP_OFFLOAD_CONT31:%.*]]
+// CHECK3:       omp_offload.failed30:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l62(i32 [[TMP93]], i32 [[TMP95]], i64* @Gblb, i32 [[TMP97]], i32 [[TMP99]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT31]]
+// CHECK3:       omp_offload.cont31:
+// CHECK3-NEXT:    [[TMP142:%.*]] = load i32*, i32** @Gblc, align 4
+// CHECK3-NEXT:    store i32* [[TMP142]], i32** [[TMP]], align 4
+// CHECK3-NEXT:    [[TMP143:%.*]] = load i32, i32* @Gbla, align 4
+// CHECK3-NEXT:    store i32 [[TMP143]], i32* [[GBLC_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP144:%.*]] = load i32, i32* [[GBLC_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP145:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK3-NEXT:    store i32 [[TMP145]], i32* [[COMP_CASTED32]], align 4
+// CHECK3-NEXT:    [[TMP146:%.*]] = load i32, i32* [[COMP_CASTED32]], align 4
+// CHECK3-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP144]], i32* [[TMP148]], align 4
+// CHECK3-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP150:%.*]] = bitcast i8** [[TMP149]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP144]], i32* [[TMP150]], align 4
+// CHECK3-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP151]], align 4
+// CHECK3-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP153:%.*]] = bitcast i8** [[TMP152]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP146]], i32* [[TMP153]], align 4
+// CHECK3-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP155:%.*]] = bitcast i8** [[TMP154]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP146]], i32* [[TMP155]], align 4
+// CHECK3-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS35]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP156]], align 4
+// CHECK3-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS33]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS34]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP159:%.*]] = load i32*, i32** @Gblc, align 4
+// CHECK3-NEXT:    store i32* [[TMP159]], i32** [[_TMP36]], align 4
+// CHECK3-NEXT:    [[TMP160:%.*]] = load i32, i32* @Gbla, align 4
+// CHECK3-NEXT:    [[ADD37:%.*]] = add nsw i32 [[TMP160]], 1
+// CHECK3-NEXT:    [[TMP161:%.*]] = load i32*, i32** @Gblc, align 4
+// CHECK3-NEXT:    store i32* [[TMP161]], i32** [[_TMP38]], align 4
+// CHECK3-NEXT:    [[TMP162:%.*]] = load i32, i32* @Gbla, align 4
+// CHECK3-NEXT:    [[ADD39:%.*]] = add nsw i32 [[TMP162]], 2
+// CHECK3-NEXT:    [[KERNEL_ARGS40:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP163:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS40]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP163]], align 4
+// CHECK3-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS40]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP164]], align 4
+// CHECK3-NEXT:    [[TMP165:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS40]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP157]], i8*** [[TMP165]], align 4
+// CHECK3-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS40]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP158]], i8*** [[TMP166]], align 4
+// CHECK3-NEXT:    [[TMP167:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS40]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP167]], align 4
+// CHECK3-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS40]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP168]], align 4
+// CHECK3-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS40]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP169]], align 4
+// CHECK3-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS40]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP170]], align 4
+// CHECK3-NEXT:    [[TMP171:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[ADD37]], i32 [[ADD39]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l71.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS40]])
+// CHECK3-NEXT:    [[TMP172:%.*]] = icmp ne i32 [[TMP171]], 0
+// CHECK3-NEXT:    br i1 [[TMP172]], label [[OMP_OFFLOAD_FAILED41:%.*]], label [[OMP_OFFLOAD_CONT42:%.*]]
+// CHECK3:       omp_offload.failed41:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l71(i32 [[TMP144]], i32 [[TMP146]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT42]]
+// CHECK3:       omp_offload.cont42:
+// CHECK3-NEXT:    [[TMP173:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP173]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z27teams_argument_global_locali_l31
@@ -1310,58 +1514,92 @@ void foo() {
 // CHECK9-NEXT:    [[TMP20:%.*]] = load float, float* [[B]], align 4
 // CHECK9-NEXT:    [[CONV1:%.*]] = fptosi float [[TMP20]] to i64
 // CHECK9-NEXT:    [[TMP21:%.*]] = trunc i64 [[CONV1]] to i32
-// CHECK9-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l116.region_id, i32 3, i8** [[TMP17]], i8** [[TMP18]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP19]], i32 [[TMP21]])
-// CHECK9-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK9-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP17]], i8*** [[TMP24]], align 8
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP18]], i8*** [[TMP25]], align 8
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 [[TMP19]], i32 [[TMP21]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l116(%struct.SS* @Gbla, %struct.SS* [[LA]], i64 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP24:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[COMP]], align 4
 // CHECK9-NEXT:    [[CONV3:%.*]] = bitcast i64* [[COMP_CASTED2]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP24]], i32* [[CONV3]], align 4
-// CHECK9-NEXT:    [[TMP25:%.*]] = load i64, i64* [[COMP_CASTED2]], align 8
-// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to %struct.SS.0**
-// CHECK9-NEXT:    store %struct.SS.0* [[LB]], %struct.SS.0** [[TMP27]], align 8
-// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to %struct.SS.0**
-// CHECK9-NEXT:    store %struct.SS.0* [[LB]], %struct.SS.0** [[TMP29]], align 8
-// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP30]], align 8
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP32:%.*]] = bitcast i8** [[TMP31]] to %struct.SS.0**
-// CHECK9-NEXT:    store %struct.SS.0* @Gblb, %struct.SS.0** [[TMP32]], align 8
-// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP34:%.*]] = bitcast i8** [[TMP33]] to %struct.SS.0**
-// CHECK9-NEXT:    store %struct.SS.0* @Gblb, %struct.SS.0** [[TMP34]], align 8
-// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP35]], align 8
-// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP25]], i64* [[TMP37]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP25]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 [[TMP32]], i32* [[CONV3]], align 4
+// CHECK9-NEXT:    [[TMP33:%.*]] = load i64, i64* [[COMP_CASTED2]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS.0**
+// CHECK9-NEXT:    store %struct.SS.0* [[LB]], %struct.SS.0** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS.0**
+// CHECK9-NEXT:    store %struct.SS.0* [[LB]], %struct.SS.0** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to %struct.SS.0**
+// CHECK9-NEXT:    store %struct.SS.0* @Gblb, %struct.SS.0** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to %struct.SS.0**
+// CHECK9-NEXT:    store %struct.SS.0* @Gblb, %struct.SS.0** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
 // CHECK9-NEXT:    [[B7:%.*]] = getelementptr inbounds [[STRUCT_SS_0]], %struct.SS.0* [[LB]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP43:%.*]] = load float, float* [[B7]], align 8
-// CHECK9-NEXT:    [[CONV8:%.*]] = fptosi float [[TMP43]] to i64
-// CHECK9-NEXT:    [[TMP44:%.*]] = trunc i64 [[CONV8]] to i32
-// CHECK9-NEXT:    [[TMP45:%.*]] = load i64, i64* getelementptr inbounds ([[STRUCT_SS_0]], %struct.SS.0* @Gblb, i32 0, i32 0), align 8
-// CHECK9-NEXT:    [[TMP46:%.*]] = trunc i64 [[TMP45]] to i32
-// CHECK9-NEXT:    [[TMP47:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l125.region_id, i32 3, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 [[TMP44]], i32 [[TMP46]])
-// CHECK9-NEXT:    [[TMP48:%.*]] = icmp ne i32 [[TMP47]], 0
-// CHECK9-NEXT:    br i1 [[TMP48]], label [[OMP_OFFLOAD_FAILED9:%.*]], label [[OMP_OFFLOAD_CONT10:%.*]]
-// CHECK9:       omp_offload.failed9:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l125(%struct.SS.0* [[LB]], %struct.SS.0* @Gblb, i64 [[TMP25]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT10]]
-// CHECK9:       omp_offload.cont10:
-// CHECK9-NEXT:    [[TMP49:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP49]]
+// CHECK9-NEXT:    [[TMP51:%.*]] = load float, float* [[B7]], align 8
+// CHECK9-NEXT:    [[CONV8:%.*]] = fptosi float [[TMP51]] to i64
+// CHECK9-NEXT:    [[TMP52:%.*]] = trunc i64 [[CONV8]] to i32
+// CHECK9-NEXT:    [[TMP53:%.*]] = load i64, i64* getelementptr inbounds ([[STRUCT_SS_0]], %struct.SS.0* @Gblb, i32 0, i32 0), align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = trunc i64 [[TMP53]] to i32
+// CHECK9-NEXT:    [[KERNEL_ARGS9:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP55]], align 4
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP56]], align 4
+// CHECK9-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP49]], i8*** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP50]], i8*** [[TMP58]], align 8
+// CHECK9-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP59]], align 8
+// CHECK9-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP60]], align 8
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP61]], align 8
+// CHECK9-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP62]], align 8
+// CHECK9-NEXT:    [[TMP63:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP52]], i32 [[TMP54]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l125.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS9]])
+// CHECK9-NEXT:    [[TMP64:%.*]] = icmp ne i32 [[TMP63]], 0
+// CHECK9-NEXT:    br i1 [[TMP64]], label [[OMP_OFFLOAD_FAILED10:%.*]], label [[OMP_OFFLOAD_CONT11:%.*]]
+// CHECK9:       omp_offload.failed10:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l125(%struct.SS.0* [[LB]], %struct.SS.0* @Gblb, i64 [[TMP33]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT11]]
+// CHECK9:       omp_offload.cont11:
+// CHECK9-NEXT:    [[TMP65:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP65]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l116
@@ -1501,57 +1739,91 @@ void foo() {
 // CHECK11-NEXT:    [[TMP20:%.*]] = load float, float* [[B]], align 4
 // CHECK11-NEXT:    [[CONV:%.*]] = fptosi float [[TMP20]] to i64
 // CHECK11-NEXT:    [[TMP21:%.*]] = trunc i64 [[CONV]] to i32
-// CHECK11-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l116.region_id, i32 3, i8** [[TMP17]], i8** [[TMP18]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP19]], i32 [[TMP21]])
-// CHECK11-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK11-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP17]], i8*** [[TMP24]], align 4
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP18]], i8*** [[TMP25]], align 4
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 [[TMP19]], i32 [[TMP21]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK11-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l116(%struct.SS* @Gbla, %struct.SS* [[LA]], i32 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP24:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK11-NEXT:    store i32 [[TMP24]], i32* [[COMP_CASTED1]], align 4
-// CHECK11-NEXT:    [[TMP25:%.*]] = load i32, i32* [[COMP_CASTED1]], align 4
-// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to %struct.SS.0**
-// CHECK11-NEXT:    store %struct.SS.0* [[LB]], %struct.SS.0** [[TMP27]], align 4
-// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP29:%.*]] = bitcast i8** [[TMP28]] to %struct.SS.0**
-// CHECK11-NEXT:    store %struct.SS.0* [[LB]], %struct.SS.0** [[TMP29]], align 4
-// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP30]], align 4
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP32:%.*]] = bitcast i8** [[TMP31]] to %struct.SS.0**
-// CHECK11-NEXT:    store %struct.SS.0* @Gblb, %struct.SS.0** [[TMP32]], align 4
-// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP34:%.*]] = bitcast i8** [[TMP33]] to %struct.SS.0**
-// CHECK11-NEXT:    store %struct.SS.0* @Gblb, %struct.SS.0** [[TMP34]], align 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP35]], align 4
-// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP25]], i32* [[TMP37]], align 4
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP25]], i32* [[TMP39]], align 4
-// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP40]], align 4
-// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK11-NEXT:    store i32 [[TMP32]], i32* [[COMP_CASTED1]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* [[COMP_CASTED1]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS.0**
+// CHECK11-NEXT:    store %struct.SS.0* [[LB]], %struct.SS.0** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS.0**
+// CHECK11-NEXT:    store %struct.SS.0* [[LB]], %struct.SS.0** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to %struct.SS.0**
+// CHECK11-NEXT:    store %struct.SS.0* @Gblb, %struct.SS.0** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to %struct.SS.0**
+// CHECK11-NEXT:    store %struct.SS.0* @Gblb, %struct.SS.0** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP45]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
 // CHECK11-NEXT:    [[B5:%.*]] = getelementptr inbounds [[STRUCT_SS_0]], %struct.SS.0* [[LB]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP43:%.*]] = load float, float* [[B5]], align 4
-// CHECK11-NEXT:    [[CONV6:%.*]] = fptosi float [[TMP43]] to i64
-// CHECK11-NEXT:    [[TMP44:%.*]] = trunc i64 [[CONV6]] to i32
-// CHECK11-NEXT:    [[TMP45:%.*]] = load i64, i64* getelementptr inbounds ([[STRUCT_SS_0]], %struct.SS.0* @Gblb, i32 0, i32 0), align 4
-// CHECK11-NEXT:    [[TMP46:%.*]] = trunc i64 [[TMP45]] to i32
-// CHECK11-NEXT:    [[TMP47:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l125.region_id, i32 3, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 [[TMP44]], i32 [[TMP46]])
-// CHECK11-NEXT:    [[TMP48:%.*]] = icmp ne i32 [[TMP47]], 0
-// CHECK11-NEXT:    br i1 [[TMP48]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK11:       omp_offload.failed7:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l125(%struct.SS.0* [[LB]], %struct.SS.0* @Gblb, i32 [[TMP25]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK11:       omp_offload.cont8:
-// CHECK11-NEXT:    [[TMP49:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP49]]
+// CHECK11-NEXT:    [[TMP51:%.*]] = load float, float* [[B5]], align 4
+// CHECK11-NEXT:    [[CONV6:%.*]] = fptosi float [[TMP51]] to i64
+// CHECK11-NEXT:    [[TMP52:%.*]] = trunc i64 [[CONV6]] to i32
+// CHECK11-NEXT:    [[TMP53:%.*]] = load i64, i64* getelementptr inbounds ([[STRUCT_SS_0]], %struct.SS.0* @Gblb, i32 0, i32 0), align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = trunc i64 [[TMP53]] to i32
+// CHECK11-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP55]], align 4
+// CHECK11-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP56]], align 4
+// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP49]], i8*** [[TMP57]], align 4
+// CHECK11-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP50]], i8*** [[TMP58]], align 4
+// CHECK11-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP59]], align 4
+// CHECK11-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP60]], align 4
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP61]], align 4
+// CHECK11-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP62]], align 4
+// CHECK11-NEXT:    [[TMP63:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 [[TMP52]], i32 [[TMP54]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l125.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK11-NEXT:    [[TMP64:%.*]] = icmp ne i32 [[TMP63]], 0
+// CHECK11-NEXT:    br i1 [[TMP64]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK11:       omp_offload.failed8:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l125(%struct.SS.0* [[LB]], %struct.SS.0* @Gblb, i32 [[TMP33]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK11:       omp_offload.cont9:
+// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP65]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z18teams_template_argv_l116
@@ -1689,49 +1961,83 @@ void foo() {
 // CHECK17-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP14:%.*]] = load i32, i32* [[A2]], align 4
-// CHECK17-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l161.region_id, i32 2, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP14]], i32 123)
-// CHECK17-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
-// CHECK17-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP15]], align 4
+// CHECK17-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 2, i32* [[TMP16]], align 4
+// CHECK17-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP12]], i8*** [[TMP17]], align 8
+// CHECK17-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP13]], i8*** [[TMP18]], align 8
+// CHECK17-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP19]], align 8
+// CHECK17-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP20]], align 8
+// CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP21]], align 8
+// CHECK17-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP22]], align 8
+// CHECK17-NEXT:    [[TMP23:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 [[TMP14]], i32 123, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l161.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP24:%.*]] = icmp ne i32 [[TMP23]], 0
+// CHECK17-NEXT:    br i1 [[TMP24]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l161(%struct.SS* [[THIS1]], i64 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
-// CHECK17-NEXT:    [[TMP17:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK17-NEXT:    [[TMP25:%.*]] = load i32, i32* [[COMP]], align 4
 // CHECK17-NEXT:    [[CONV4:%.*]] = bitcast i64* [[COMP_CASTED3]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP17]], i32* [[CONV4]], align 4
-// CHECK17-NEXT:    [[TMP18:%.*]] = load i64, i64* [[COMP_CASTED3]], align 8
+// CHECK17-NEXT:    store i32 [[TMP25]], i32* [[CONV4]], align 4
+// CHECK17-NEXT:    [[TMP26:%.*]] = load i64, i64* [[COMP_CASTED3]], align 8
 // CHECK17-NEXT:    [[B:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to %struct.SS**
-// CHECK17-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP20]], align 8
-// CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to float**
-// CHECK17-NEXT:    store float* [[B]], float** [[TMP22]], align 8
-// CHECK17-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP23]], align 8
-// CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP18]], i64* [[TMP25]], align 8
-// CHECK17-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP18]], i64* [[TMP27]], align 8
-// CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP28]], align 8
-// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
+// CHECK17-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 8
+// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to float**
+// CHECK17-NEXT:    store float* [[B]], float** [[TMP30]], align 8
+// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP31]], align 8
+// CHECK17-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP33:%.*]] = bitcast i8** [[TMP32]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP26]], i64* [[TMP33]], align 8
+// CHECK17-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP26]], i64* [[TMP35]], align 8
+// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP36]], align 8
+// CHECK17-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
 // CHECK17-NEXT:    [[B8:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP31:%.*]] = load float, float* [[B8]], align 4
-// CHECK17-NEXT:    [[CONV9:%.*]] = fptosi float [[TMP31]] to i32
+// CHECK17-NEXT:    [[TMP39:%.*]] = load float, float* [[B8]], align 4
+// CHECK17-NEXT:    [[CONV9:%.*]] = fptosi float [[TMP39]] to i32
 // CHECK17-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV9]], 123
-// CHECK17-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l169.region_id, i32 2, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 456, i32 [[ADD]])
-// CHECK17-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK17-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED10:%.*]], label [[OMP_OFFLOAD_CONT11:%.*]]
-// CHECK17:       omp_offload.failed10:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l169(%struct.SS* [[THIS1]], i64 [[TMP18]]) #[[ATTR2]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT11]]
-// CHECK17:       omp_offload.cont11:
-// CHECK17-NEXT:    [[TMP34:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK17-NEXT:    ret i32 [[TMP34]]
+// CHECK17-NEXT:    [[KERNEL_ARGS10:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 2, i32* [[TMP41]], align 4
+// CHECK17-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 8
+// CHECK17-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 8
+// CHECK17-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP44]], align 8
+// CHECK17-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP46]], align 8
+// CHECK17-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK17-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 456, i32 [[ADD]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l169.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS10]])
+// CHECK17-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK17-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
+// CHECK17:       omp_offload.failed11:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l169(%struct.SS* [[THIS1]], i64 [[TMP26]]) #[[ATTR2]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
+// CHECK17:       omp_offload.cont12:
+// CHECK17-NEXT:    [[TMP50:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK17-NEXT:    ret i32 [[TMP50]]
 //
 //
 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l161
@@ -1857,48 +2163,82 @@ void foo() {
 // CHECK19-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP14:%.*]] = load i32, i32* [[A2]], align 4
-// CHECK19-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l161.region_id, i32 2, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP14]], i32 123)
-// CHECK19-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
-// CHECK19-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP15]], align 4
+// CHECK19-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 2, i32* [[TMP16]], align 4
+// CHECK19-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP12]], i8*** [[TMP17]], align 4
+// CHECK19-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP13]], i8*** [[TMP18]], align 4
+// CHECK19-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP19]], align 4
+// CHECK19-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP20]], align 4
+// CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP21]], align 4
+// CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP22]], align 4
+// CHECK19-NEXT:    [[TMP23:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 [[TMP14]], i32 123, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l161.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP24:%.*]] = icmp ne i32 [[TMP23]], 0
+// CHECK19-NEXT:    br i1 [[TMP24]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l161(%struct.SS* [[THIS1]], i32 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
-// CHECK19-NEXT:    [[TMP17:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK19-NEXT:    store i32 [[TMP17]], i32* [[COMP_CASTED3]], align 4
-// CHECK19-NEXT:    [[TMP18:%.*]] = load i32, i32* [[COMP_CASTED3]], align 4
+// CHECK19-NEXT:    [[TMP25:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK19-NEXT:    store i32 [[TMP25]], i32* [[COMP_CASTED3]], align 4
+// CHECK19-NEXT:    [[TMP26:%.*]] = load i32, i32* [[COMP_CASTED3]], align 4
 // CHECK19-NEXT:    [[B:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to %struct.SS**
-// CHECK19-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP20]], align 4
-// CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP22:%.*]] = bitcast i8** [[TMP21]] to float**
-// CHECK19-NEXT:    store float* [[B]], float** [[TMP22]], align 4
-// CHECK19-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP23]], align 4
-// CHECK19-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP18]], i32* [[TMP25]], align 4
-// CHECK19-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP27:%.*]] = bitcast i8** [[TMP26]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP18]], i32* [[TMP27]], align 4
-// CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP28]], align 4
-// CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
+// CHECK19-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 4
+// CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to float**
+// CHECK19-NEXT:    store float* [[B]], float** [[TMP30]], align 4
+// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP31]], align 4
+// CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP33:%.*]] = bitcast i8** [[TMP32]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP26]], i32* [[TMP33]], align 4
+// CHECK19-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP26]], i32* [[TMP35]], align 4
+// CHECK19-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP36]], align 4
+// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
 // CHECK19-NEXT:    [[B7:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP31:%.*]] = load float, float* [[B7]], align 4
-// CHECK19-NEXT:    [[CONV:%.*]] = fptosi float [[TMP31]] to i32
+// CHECK19-NEXT:    [[TMP39:%.*]] = load float, float* [[B7]], align 4
+// CHECK19-NEXT:    [[CONV:%.*]] = fptosi float [[TMP39]] to i32
 // CHECK19-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV]], 123
-// CHECK19-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l169.region_id, i32 2, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 456, i32 [[ADD]])
-// CHECK19-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK19-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
-// CHECK19:       omp_offload.failed8:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l169(%struct.SS* [[THIS1]], i32 [[TMP18]]) #[[ATTR2]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
-// CHECK19:       omp_offload.cont9:
-// CHECK19-NEXT:    [[TMP34:%.*]] = load i32, i32* [[COMP]], align 4
-// CHECK19-NEXT:    ret i32 [[TMP34]]
+// CHECK19-NEXT:    [[KERNEL_ARGS8:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 2, i32* [[TMP41]], align 4
+// CHECK19-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP37]], i8*** [[TMP42]], align 4
+// CHECK19-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP38]], i8*** [[TMP43]], align 4
+// CHECK19-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP44]], align 4
+// CHECK19-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK19-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP46]], align 4
+// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK19-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 456, i32 [[ADD]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l169.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS8]])
+// CHECK19-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK19-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED9:%.*]], label [[OMP_OFFLOAD_CONT10:%.*]]
+// CHECK19:       omp_offload.failed9:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l169(%struct.SS* [[THIS1]], i32 [[TMP26]]) #[[ATTR2]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT10]]
+// CHECK19:       omp_offload.cont10:
+// CHECK19-NEXT:    [[TMP50:%.*]] = load i32, i32* [[COMP]], align 4
+// CHECK19-NEXT:    ret i32 [[TMP50]]
 //
 //
 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l161

diff --git a/clang/test/OpenMP/teams_distribute_codegen.cpp b/clang/test/OpenMP/teams_distribute_codegen.cpp
index 6cd94ad..52a9e6c 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_codegen.cpp
@@ -261,55 +261,89 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP33]], 1
 // CHECK1-NEXT:    [[TMP34:%.*]] = zext i32 [[ADD]] to i64
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP34]])
-// CHECK1-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP29]], i32 [[TMP30]])
-// CHECK1-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
-// CHECK1-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP35]], align 4
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP36]], align 4
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP41]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK1-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP29]], i32 [[TMP30]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
+// CHECK1-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29(i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], [100 x i32]* @a) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP37:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[TMP45:%.*]] = load i32, i32* [[N_ADDR]], align 4
 // CHECK1-NEXT:    [[CONV7:%.*]] = bitcast i64* [[N_CASTED6]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP37]], i32* [[CONV7]], align 4
-// CHECK1-NEXT:    [[TMP38:%.*]] = load i64, i64* [[N_CASTED6]], align 8
-// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP38]], i64* [[TMP40]], align 8
-// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP38]], i64* [[TMP42]], align 8
-// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP43]], align 8
-// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to [100 x i32]**
-// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP45]], align 8
-// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to [100 x i32]**
-// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP47]], align 8
-// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP48]], align 8
-// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP51:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    store i32 [[TMP51]], i32* [[DOTCAPTURE_EXPR_12]], align 4
-// CHECK1-NEXT:    [[TMP52:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_12]], align 4
-// CHECK1-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[TMP52]], 0
+// CHECK1-NEXT:    store i32 [[TMP45]], i32* [[CONV7]], align 4
+// CHECK1-NEXT:    [[TMP46:%.*]] = load i64, i64* [[N_CASTED6]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP46]], i64* [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP46]], i64* [[TMP50]], align 8
+// CHECK1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP51]], align 8
+// CHECK1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to [100 x i32]**
+// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP53]], align 8
+// CHECK1-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to [100 x i32]**
+// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP55]], align 8
+// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP56]], align 8
+// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP59:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    store i32 [[TMP59]], i32* [[DOTCAPTURE_EXPR_12]], align 4
+// CHECK1-NEXT:    [[TMP60:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_12]], align 4
+// CHECK1-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[TMP60]], 0
 // CHECK1-NEXT:    [[DIV15:%.*]] = sdiv i32 [[SUB14]], 1
 // CHECK1-NEXT:    [[SUB16:%.*]] = sub nsw i32 [[DIV15]], 1
 // CHECK1-NEXT:    store i32 [[SUB16]], i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK1-NEXT:    [[TMP53:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK1-NEXT:    [[ADD17:%.*]] = add nsw i32 [[TMP53]], 1
-// CHECK1-NEXT:    [[TMP54:%.*]] = zext i32 [[ADD17]] to i64
-// CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP54]])
-// CHECK1-NEXT:    [[TMP55:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35.region_id, i32 2, i8** [[TMP49]], i8** [[TMP50]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP56:%.*]] = icmp ne i32 [[TMP55]], 0
-// CHECK1-NEXT:    br i1 [[TMP56]], label [[OMP_OFFLOAD_FAILED18:%.*]], label [[OMP_OFFLOAD_CONT19:%.*]]
-// CHECK1:       omp_offload.failed18:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35(i64 [[TMP38]], [100 x i32]* @a) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT19]]
-// CHECK1:       omp_offload.cont19:
-// CHECK1-NEXT:    [[TMP57:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i64 0, i64 0), align 4
-// CHECK1-NEXT:    ret i32 [[TMP57]]
+// CHECK1-NEXT:    [[TMP61:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
+// CHECK1-NEXT:    [[ADD17:%.*]] = add nsw i32 [[TMP61]], 1
+// CHECK1-NEXT:    [[TMP62:%.*]] = zext i32 [[ADD17]] to i64
+// CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP62]])
+// CHECK1-NEXT:    [[KERNEL_ARGS18:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP63]], align 4
+// CHECK1-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP64]], align 4
+// CHECK1-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP57]], i8*** [[TMP65]], align 8
+// CHECK1-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP58]], i8*** [[TMP66]], align 8
+// CHECK1-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP67]], align 8
+// CHECK1-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP68]], align 8
+// CHECK1-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP69]], align 8
+// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP70]], align 8
+// CHECK1-NEXT:    [[TMP71:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]])
+// CHECK1-NEXT:    [[TMP72:%.*]] = icmp ne i32 [[TMP71]], 0
+// CHECK1-NEXT:    br i1 [[TMP72]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
+// CHECK1:       omp_offload.failed19:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35(i64 [[TMP46]], [100 x i32]* @a) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
+// CHECK1:       omp_offload.cont20:
+// CHECK1-NEXT:    [[TMP73:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i64 0, i64 0), align 4
+// CHECK1-NEXT:    ret i32 [[TMP73]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29
@@ -622,54 +656,88 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP33]], 1
 // CHECK3-NEXT:    [[TMP34:%.*]] = zext i32 [[ADD]] to i64
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP34]])
-// CHECK3-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP29]], i32 [[TMP30]])
-// CHECK3-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
-// CHECK3-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP39]], align 4
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP29]], i32 [[TMP30]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
+// CHECK3-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29(i32 [[TMP2]], i32 [[TMP4]], i32 [[TMP6]], [100 x i32]* @a) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP37:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    store i32 [[TMP37]], i32* [[N_CASTED4]], align 4
-// CHECK3-NEXT:    [[TMP38:%.*]] = load i32, i32* [[N_CASTED4]], align 4
-// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP38]], i32* [[TMP40]], align 4
-// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP38]], i32* [[TMP42]], align 4
-// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP43]], align 4
-// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to [100 x i32]**
-// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP45]], align 4
-// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to [100 x i32]**
-// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP47]], align 4
-// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP48]], align 4
-// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP51:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    store i32 [[TMP51]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK3-NEXT:    [[TMP52:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK3-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP52]], 0
+// CHECK3-NEXT:    [[TMP45:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    store i32 [[TMP45]], i32* [[N_CASTED4]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = load i32, i32* [[N_CASTED4]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP46]], i32* [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP46]], i32* [[TMP50]], align 4
+// CHECK3-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP51]], align 4
+// CHECK3-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to [100 x i32]**
+// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP53]], align 4
+// CHECK3-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to [100 x i32]**
+// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP55]], align 4
+// CHECK3-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP56]], align 4
+// CHECK3-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP59:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    store i32 [[TMP59]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK3-NEXT:    [[TMP60:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK3-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP60]], 0
 // CHECK3-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK3-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK3-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK3-NEXT:    [[TMP53:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK3-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP53]], 1
-// CHECK3-NEXT:    [[TMP54:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP54]])
-// CHECK3-NEXT:    [[TMP55:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35.region_id, i32 2, i8** [[TMP49]], i8** [[TMP50]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP56:%.*]] = icmp ne i32 [[TMP55]], 0
-// CHECK3-NEXT:    br i1 [[TMP56]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK3:       omp_offload.failed15:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35(i32 [[TMP38]], [100 x i32]* @a) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK3:       omp_offload.cont16:
-// CHECK3-NEXT:    [[TMP57:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i32 0, i32 0), align 4
-// CHECK3-NEXT:    ret i32 [[TMP57]]
+// CHECK3-NEXT:    [[TMP61:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK3-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP61]], 1
+// CHECK3-NEXT:    [[TMP62:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP62]])
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP63]], align 4
+// CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP64]], align 4
+// CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP57]], i8*** [[TMP65]], align 4
+// CHECK3-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP58]], i8*** [[TMP66]], align 4
+// CHECK3-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP67]], align 4
+// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP68]], align 4
+// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP70]], align 4
+// CHECK3-NEXT:    [[TMP71:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP72:%.*]] = icmp ne i32 [[TMP71]], 0
+// CHECK3-NEXT:    br i1 [[TMP72]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35(i32 [[TMP46]], [100 x i32]* @a) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[TMP73:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i32 0, i32 0), align 4
+// CHECK3-NEXT:    ret i32 [[TMP73]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29
@@ -961,18 +1029,35 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK9-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK9-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 0
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP33]])
-// CHECK9-NEXT:    ret i32 [[TMP32]]
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP41]])
+// CHECK9-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73
@@ -1160,18 +1245,35 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK11-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK11-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK11-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK11-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i32 0
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP33]])
-// CHECK11-NEXT:    ret i32 [[TMP32]]
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP41]])
+// CHECK11-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73
@@ -1320,17 +1422,34 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK17-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK17-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK17-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK17-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK17-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK17-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK17-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK17-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK17-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
 // CHECK17-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK17-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A2]], i64 0, i64 0
-// CHECK17-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK17-NEXT:    ret i32 [[TMP9]]
+// CHECK17-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK17-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109
@@ -1448,17 +1567,34 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK19-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK19-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK19-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK19-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK19-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK19-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK19-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK19-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK19-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
 // CHECK19-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK19-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A2]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK19-NEXT:    ret i32 [[TMP9]]
+// CHECK19-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK19-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109
@@ -1619,20 +1755,37 @@ int main (int argc, char **argv) {
 // CHECK25-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK25-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK25-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK25-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK25-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK25-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK25-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK25-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK25-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK25-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK25-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK25-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK25-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK25-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK25-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK25-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK25-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK25-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK25-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK25-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK25-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK25-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK25-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK25-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK25-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK25-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK25:       omp_offload.failed:
 // CHECK25-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK25-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK25:       omp_offload.cont:
-// CHECK25-NEXT:    [[TMP32:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK25-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP32]])
+// CHECK25-NEXT:    [[TMP40:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK25-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP40]])
 // CHECK25-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK25-NEXT:    [[TMP33:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK25-NEXT:    call void @llvm.stackrestore(i8* [[TMP33]])
-// CHECK25-NEXT:    [[TMP34:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK25-NEXT:    ret i32 [[TMP34]]
+// CHECK25-NEXT:    [[TMP41:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK25-NEXT:    call void @llvm.stackrestore(i8* [[TMP41]])
+// CHECK25-NEXT:    [[TMP42:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK25-NEXT:    ret i32 [[TMP42]]
 //
 //
 // CHECK25-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162
@@ -1799,9 +1952,26 @@ int main (int argc, char **argv) {
 // CHECK25-NEXT:    [[TMP21:%.*]] = load i32, i32* [[TE]], align 4
 // CHECK25-NEXT:    [[TMP22:%.*]] = load i32, i32* [[TH]], align 4
 // CHECK25-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK25-NEXT:    [[TMP23:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l151.region_id, i32 3, i8** [[TMP19]], i8** [[TMP20]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 [[TMP21]], i32 [[TMP22]])
-// CHECK25-NEXT:    [[TMP24:%.*]] = icmp ne i32 [[TMP23]], 0
-// CHECK25-NEXT:    br i1 [[TMP24]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK25-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK25-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK25-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK25-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK25-NEXT:    store i32 3, i32* [[TMP24]], align 4
+// CHECK25-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK25-NEXT:    store i8** [[TMP19]], i8*** [[TMP25]], align 8
+// CHECK25-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK25-NEXT:    store i8** [[TMP20]], i8*** [[TMP26]], align 8
+// CHECK25-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK25-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK25-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK25-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK25-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK25-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK25-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK25-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK25-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP21]], i32 [[TMP22]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l151.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK25-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
+// CHECK25-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK25:       omp_offload.failed:
 // CHECK25-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l151(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK25-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1976,20 +2146,37 @@ int main (int argc, char **argv) {
 // CHECK27-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK27-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK27-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK27-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK27-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK27-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK27-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK27-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK27-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK27-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK27-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK27-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK27-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK27-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK27-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK27-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK27-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK27-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK27-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK27-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK27-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK27-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK27-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK27-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK27-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK27-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK27:       omp_offload.failed:
 // CHECK27-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK27-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK27:       omp_offload.cont:
-// CHECK27-NEXT:    [[TMP32:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK27-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP32]])
+// CHECK27-NEXT:    [[TMP40:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK27-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP40]])
 // CHECK27-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK27-NEXT:    [[TMP33:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK27-NEXT:    call void @llvm.stackrestore(i8* [[TMP33]])
-// CHECK27-NEXT:    [[TMP34:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK27-NEXT:    ret i32 [[TMP34]]
+// CHECK27-NEXT:    [[TMP41:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK27-NEXT:    call void @llvm.stackrestore(i8* [[TMP41]])
+// CHECK27-NEXT:    [[TMP42:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK27-NEXT:    ret i32 [[TMP42]]
 //
 //
 // CHECK27-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162
@@ -2152,9 +2339,26 @@ int main (int argc, char **argv) {
 // CHECK27-NEXT:    [[TMP21:%.*]] = load i32, i32* [[TE]], align 4
 // CHECK27-NEXT:    [[TMP22:%.*]] = load i32, i32* [[TH]], align 4
 // CHECK27-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK27-NEXT:    [[TMP23:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l151.region_id, i32 3, i8** [[TMP19]], i8** [[TMP20]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 [[TMP21]], i32 [[TMP22]])
-// CHECK27-NEXT:    [[TMP24:%.*]] = icmp ne i32 [[TMP23]], 0
-// CHECK27-NEXT:    br i1 [[TMP24]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK27-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK27-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK27-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK27-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK27-NEXT:    store i32 3, i32* [[TMP24]], align 4
+// CHECK27-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK27-NEXT:    store i8** [[TMP19]], i8*** [[TMP25]], align 4
+// CHECK27-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK27-NEXT:    store i8** [[TMP20]], i8*** [[TMP26]], align 4
+// CHECK27-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK27-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK27-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK27-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK27-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK27-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK27-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK27-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK27-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP21]], i32 [[TMP22]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l151.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK27-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
+// CHECK27-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK27:       omp_offload.failed:
 // CHECK27-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l151(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK27-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_collapse_codegen.cpp b/clang/test/OpenMP/teams_distribute_collapse_codegen.cpp
index 1bda312..f548776 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_collapse_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_collapse_codegen.cpp
@@ -128,9 +128,26 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 56088)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -138,8 +155,8 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i64 0, i64 0
 // CHECK1-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP9]]
+// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -272,9 +289,26 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 56088)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -282,8 +316,8 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP9]]
+// CHECK3-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -495,20 +529,37 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP46:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_4]], align 8
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP46]], 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK9-NEXT:    [[TMP47:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l82.region_id, i32 5, i8** [[TMP39]], i8** [[TMP40]], i64* [[TMP41]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP48:%.*]] = icmp ne i32 [[TMP47]], 0
-// CHECK9-NEXT:    br i1 [[TMP48]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP47]], align 4
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP48]], align 4
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 8
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP40]], i8*** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP41]], i64** [[TMP51]], align 8
+// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l82.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP56:%.*]] = icmp ne i32 [[TMP55]], 0
+// CHECK9-NEXT:    br i1 [[TMP56]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l82(i64 [[TMP7]], i64 [[TMP9]], i64 [[TMP1]], i64 [[TMP3]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP49:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP49]])
+// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP57]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP50:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP50]])
-// CHECK9-NEXT:    [[TMP51:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP51]]
+// CHECK9-NEXT:    [[TMP58:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP58]])
+// CHECK9-NEXT:    [[TMP59:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP59]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l82
@@ -703,9 +754,26 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 20)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -920,20 +988,37 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP45:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_3]], align 8
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP45]], 1
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK11-NEXT:    [[TMP46:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l82.region_id, i32 5, i8** [[TMP38]], i8** [[TMP39]], i64* [[TMP40]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP47:%.*]] = icmp ne i32 [[TMP46]], 0
-// CHECK11-NEXT:    br i1 [[TMP47]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP38]], i8*** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP40]], i64** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l82.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP55:%.*]] = icmp ne i32 [[TMP54]], 0
+// CHECK11-NEXT:    br i1 [[TMP55]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l82(i32 [[TMP5]], i32 [[TMP7]], i32 [[TMP0]], i32 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP48:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP48]])
+// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP56]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP49]])
-// CHECK11-NEXT:    [[TMP50:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP50]]
+// CHECK11-NEXT:    [[TMP57:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP57]])
+// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP58]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l82
@@ -1124,9 +1209,26 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 20)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l68([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_dist_schedule_codegen.cpp b/clang/test/OpenMP/teams_distribute_dist_schedule_codegen.cpp
index e733381..be67c9f 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_dist_schedule_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_dist_schedule_codegen.cpp
@@ -151,10 +151,10 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -169,55 +169,106 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK1:       omp_offload.failed16:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK1:       omp_offload.cont15:
-// CHECK1-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP27]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK1:       omp_offload.cont17:
+// CHECK1-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i64 0, i64 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -498,10 +549,10 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -516,55 +567,106 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK3:       omp_offload.failed8:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK3:       omp_offload.cont8:
-// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK3:       omp_offload.cont9:
+// CHECK3-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK3-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK3:       omp_offload.failed14:
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK3:       omp_offload.cont15:
-// CHECK3-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP27]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -847,14 +949,14 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[_TMP9:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED18:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [3 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [3 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [3 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES23:%.*]] = alloca [3 x i64], align 8
-// CHECK9-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [3 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [3 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [3 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES24:%.*]] = alloca [3 x i64], align 8
+// CHECK9-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -912,128 +1014,179 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK9-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK9-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK9-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i64 24, i1 false)
-// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK9-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i64 24, i1 false)
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK9-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK9-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK9-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK9-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK9-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK9-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP58]])
-// CHECK9-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK9-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK9:       omp_offload.failed16:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK9:       omp_offload.cont17:
-// CHECK9-NEXT:    [[TMP61:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV19:%.*]] = bitcast i64* [[N_CASTED18]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP61]], i32* [[CONV19]], align 4
-// CHECK9-NEXT:    [[TMP62:%.*]] = load i64, i64* [[N_CASTED18]], align 8
-// CHECK9-NEXT:    [[TMP63:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP64:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES23]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP64]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.5 to i8*), i64 24, i1 false)
-// CHECK9-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP66:%.*]] = bitcast i8** [[TMP65]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP62]], i64* [[TMP66]], align 8
-// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP62]], i64* [[TMP68]], align 8
-// CHECK9-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP69]], align 8
-// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP71]], align 8
-// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP73]], align 8
-// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP74]], align 8
-// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP76]], align 8
-// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP78]], align 8
-// CHECK9-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP63]], i64* [[TMP79]], align 8
-// CHECK9-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP80]], align 8
-// CHECK9-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP84:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP84]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK9-NEXT:    [[TMP85:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK9-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP85]], 0
-// CHECK9-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
-// CHECK9-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
-// CHECK9-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK9-NEXT:    [[TMP86:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK9-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP86]], 1
-// CHECK9-NEXT:    [[TMP87:%.*]] = zext i32 [[ADD30]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP87]])
-// CHECK9-NEXT:    [[TMP88:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110.region_id, i32 3, i8** [[TMP81]], i8** [[TMP82]], i64* [[TMP83]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP89:%.*]] = icmp ne i32 [[TMP88]], 0
-// CHECK9-NEXT:    br i1 [[TMP89]], label [[OMP_OFFLOAD_FAILED31:%.*]], label [[OMP_OFFLOAD_CONT32:%.*]]
-// CHECK9:       omp_offload.failed31:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110(i64 [[TMP62]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT32]]
-// CHECK9:       omp_offload.cont32:
-// CHECK9-NEXT:    [[TMP90:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP90]])
+// CHECK9-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK9-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP66]])
+// CHECK9-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK9-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK9-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK9-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK9-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK9-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK9-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK9:       omp_offload.failed17:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK9:       omp_offload.cont18:
+// CHECK9-NEXT:    [[TMP77:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP77]], i32* [[CONV20]], align 4
+// CHECK9-NEXT:    [[TMP78:%.*]] = load i64, i64* [[N_CASTED19]], align 8
+// CHECK9-NEXT:    [[TMP79:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP80:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES24]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP80]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.5 to i8*), i64 24, i1 false)
+// CHECK9-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP78]], i64* [[TMP82]], align 8
+// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP78]], i64* [[TMP84]], align 8
+// CHECK9-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP85]], align 8
+// CHECK9-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP87]], align 8
+// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP89]], align 8
+// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP90]], align 8
+// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP92]], align 8
+// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP94]], align 8
+// CHECK9-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP79]], i64* [[TMP95]], align 8
+// CHECK9-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP96]], align 8
+// CHECK9-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP100]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK9-NEXT:    [[TMP101:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK9-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[TMP101]], 0
+// CHECK9-NEXT:    [[DIV29:%.*]] = sdiv i32 [[SUB28]], 1
+// CHECK9-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[DIV29]], 1
+// CHECK9-NEXT:    store i32 [[SUB30]], i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK9-NEXT:    [[TMP102:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK9-NEXT:    [[ADD31:%.*]] = add nsw i32 [[TMP102]], 1
+// CHECK9-NEXT:    [[TMP103:%.*]] = zext i32 [[ADD31]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP103]])
+// CHECK9-NEXT:    [[KERNEL_ARGS32:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP104]], align 4
+// CHECK9-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP105]], align 4
+// CHECK9-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP97]], i8*** [[TMP106]], align 8
+// CHECK9-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP98]], i8*** [[TMP107]], align 8
+// CHECK9-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP99]], i64** [[TMP108]], align 8
+// CHECK9-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP109]], align 8
+// CHECK9-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP110]], align 8
+// CHECK9-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP111]], align 8
+// CHECK9-NEXT:    [[TMP112:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]])
+// CHECK9-NEXT:    [[TMP113:%.*]] = icmp ne i32 [[TMP112]], 0
+// CHECK9-NEXT:    br i1 [[TMP113]], label [[OMP_OFFLOAD_FAILED33:%.*]], label [[OMP_OFFLOAD_CONT34:%.*]]
+// CHECK9:       omp_offload.failed33:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110(i64 [[TMP78]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT34]]
+// CHECK9:       omp_offload.cont34:
+// CHECK9-NEXT:    [[TMP114:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP114]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP91:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP91]])
-// CHECK9-NEXT:    [[TMP92:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP92]]
+// CHECK9-NEXT:    [[TMP115:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP115]])
+// CHECK9-NEXT:    [[TMP116:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP116]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100
@@ -1411,10 +1564,10 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[DOTOFFLOAD_PTRS2:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [1 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [1 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [1 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP1:%.*]] = bitcast i8** [[TMP0]] to [10 x i32]**
@@ -1427,49 +1580,100 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l79.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l79.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l79([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK9-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK9:       omp_offload.failed5:
+// CHECK9-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK9-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK9-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK9:       omp_offload.failed6:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK9:       omp_offload.cont6:
-// CHECK9-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP19]], align 8
-// CHECK9-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP21]], align 8
-// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK9:       omp_offload.cont7:
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK9-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK9:       omp_offload.failed11:
+// CHECK9-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK9-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK9-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK9:       omp_offload.failed13:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK9:       omp_offload.cont12:
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK9:       omp_offload.cont14:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -1753,14 +1957,14 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[_TMP8:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED17:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS18:%.*]] = alloca [3 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS19:%.*]] = alloca [3 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS20:%.*]] = alloca [3 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES21:%.*]] = alloca [3 x i64], align 4
-// CHECK11-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_23:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [3 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [3 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [3 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [3 x i64], align 4
+// CHECK11-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -1817,128 +2021,179 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK11-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK11-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK11-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK11-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK11-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i32 24, i1 false)
-// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i32 24, i1 false)
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK11-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK11-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK11-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK11-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK11-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK11-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP59]])
-// CHECK11-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK11-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK11:       omp_offload.failed15:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK11:       omp_offload.cont16:
-// CHECK11-NEXT:    [[TMP62:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP62]], i32* [[N_CASTED17]], align 4
-// CHECK11-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N_CASTED17]], align 4
-// CHECK11-NEXT:    [[TMP64:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP65:%.*]] = sext i32 [[TMP64]] to i64
-// CHECK11-NEXT:    [[TMP66:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES21]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP66]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.5 to i8*), i32 24, i1 false)
-// CHECK11-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP63]], i32* [[TMP68]], align 4
-// CHECK11-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP63]], i32* [[TMP70]], align 4
-// CHECK11-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP71]], align 4
-// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP73]], align 4
-// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP75]], align 4
-// CHECK11-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP76]], align 4
-// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP78]], align 4
-// CHECK11-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP80]], align 4
-// CHECK11-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP65]], i64* [[TMP81]], align 4
-// CHECK11-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP82]], align 4
-// CHECK11-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP86:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP86]], i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK11-NEXT:    [[TMP87:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK11-NEXT:    [[SUB25:%.*]] = sub nsw i32 [[TMP87]], 0
-// CHECK11-NEXT:    [[DIV26:%.*]] = sdiv i32 [[SUB25]], 1
-// CHECK11-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[DIV26]], 1
-// CHECK11-NEXT:    store i32 [[SUB27]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[TMP88:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[ADD28:%.*]] = add nsw i32 [[TMP88]], 1
-// CHECK11-NEXT:    [[TMP89:%.*]] = zext i32 [[ADD28]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP89]])
-// CHECK11-NEXT:    [[TMP90:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110.region_id, i32 3, i8** [[TMP83]], i8** [[TMP84]], i64* [[TMP85]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP91:%.*]] = icmp ne i32 [[TMP90]], 0
-// CHECK11-NEXT:    br i1 [[TMP91]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
-// CHECK11:       omp_offload.failed29:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110(i32 [[TMP63]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
-// CHECK11:       omp_offload.cont30:
-// CHECK11-NEXT:    [[TMP92:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP92]])
+// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK11-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP67]])
+// CHECK11-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK11-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK11-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK11-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK11-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK11-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK11-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK11:       omp_offload.failed16:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK11:       omp_offload.cont17:
+// CHECK11-NEXT:    [[TMP78:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP78]], i32* [[N_CASTED18]], align 4
+// CHECK11-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N_CASTED18]], align 4
+// CHECK11-NEXT:    [[TMP80:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP81:%.*]] = sext i32 [[TMP80]] to i64
+// CHECK11-NEXT:    [[TMP82:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP82]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.5 to i8*), i32 24, i1 false)
+// CHECK11-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP79]], i32* [[TMP84]], align 4
+// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP79]], i32* [[TMP86]], align 4
+// CHECK11-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP87]], align 4
+// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP89]], align 4
+// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP91]], align 4
+// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP92]], align 4
+// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP94]], align 4
+// CHECK11-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP96]], align 4
+// CHECK11-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP81]], i64* [[TMP97]], align 4
+// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP98]], align 4
+// CHECK11-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP102:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP102]], i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK11-NEXT:    [[TMP103:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK11-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP103]], 0
+// CHECK11-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
+// CHECK11-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
+// CHECK11-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[TMP104:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP104]], 1
+// CHECK11-NEXT:    [[TMP105:%.*]] = zext i32 [[ADD29]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP105]])
+// CHECK11-NEXT:    [[KERNEL_ARGS30:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP106]], align 4
+// CHECK11-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP107]], align 4
+// CHECK11-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP99]], i8*** [[TMP108]], align 4
+// CHECK11-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP100]], i8*** [[TMP109]], align 4
+// CHECK11-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP101]], i64** [[TMP110]], align 4
+// CHECK11-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP111]], align 4
+// CHECK11-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP112]], align 4
+// CHECK11-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP113]], align 4
+// CHECK11-NEXT:    [[TMP114:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]])
+// CHECK11-NEXT:    [[TMP115:%.*]] = icmp ne i32 [[TMP114]], 0
+// CHECK11-NEXT:    br i1 [[TMP115]], label [[OMP_OFFLOAD_FAILED31:%.*]], label [[OMP_OFFLOAD_CONT32:%.*]]
+// CHECK11:       omp_offload.failed31:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110(i32 [[TMP79]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT32]]
+// CHECK11:       omp_offload.cont32:
+// CHECK11-NEXT:    [[TMP116:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP116]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP93:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP93]])
-// CHECK11-NEXT:    [[TMP94:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP94]]
+// CHECK11-NEXT:    [[TMP117:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP117]])
+// CHECK11-NEXT:    [[TMP118:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP118]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100
@@ -2308,10 +2563,10 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[DOTOFFLOAD_PTRS2:%.*]] = alloca [1 x i8*], align 4
 // CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 4
 // CHECK11-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [1 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [1 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [1 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [1 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [1 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [1 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP1:%.*]] = bitcast i8** [[TMP0]] to [10 x i32]**
@@ -2324,49 +2579,100 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l79.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l79.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l79([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK11-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK11-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK11:       omp_offload.failed5:
+// CHECK11-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK11-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK11-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK11:       omp_offload.failed6:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK11:       omp_offload.cont6:
-// CHECK11-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP19]], align 4
-// CHECK11-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP21]], align 4
-// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK11:       omp_offload.cont7:
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK11-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK11:       omp_offload.failed11:
+// CHECK11-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK11-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK11-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK11:       omp_offload.failed13:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK11:       omp_offload.cont12:
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK11:       omp_offload.cont14:
 // CHECK11-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/teams_distribute_firstprivate_codegen.cpp b/clang/test/OpenMP/teams_distribute_firstprivate_codegen.cpp
index ae0c7e5..44f88bb 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_firstprivate_codegen.cpp
@@ -317,9 +317,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK1-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK1-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94(i64 [[TMP1]], [2 x i32]* @vec, [2 x %struct.S]* @s_arr, %struct.S* @var, i64 [[TMP3]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -593,27 +610,44 @@ int main() {
 // CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK1-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK1-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP32]]
+// CHECK1-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN2StC2Ev
@@ -1107,9 +1141,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK3-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK3-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94(i32 [[TMP1]], [2 x i32]* @vec, [2 x %struct.S]* @s_arr, %struct.S* @var, i32 [[TMP3]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1374,27 +1425,44 @@ int main() {
 // CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK3-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK3-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP32]]
+// CHECK3-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN2StC2Ev

diff --git a/clang/test/OpenMP/teams_distribute_lastprivate_codegen.cpp b/clang/test/OpenMP/teams_distribute_lastprivate_codegen.cpp
index f906e2e..cf114b4 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_lastprivate_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_lastprivate_codegen.cpp
@@ -549,9 +549,26 @@ int main() {
 // CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK9-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK9-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -559,18 +576,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done3:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP39]]
+// CHECK9-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -855,27 +872,44 @@ int main() {
 // CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK9-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK9-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP32]]
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1242,9 +1276,26 @@ int main() {
 // CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK11-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK11-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1252,18 +1303,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP39]]
+// CHECK11-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1543,27 +1594,44 @@ int main() {
 // CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK11-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK11-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP32]]
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_codegen.cpp
index 70e3870..5fc701e 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_codegen.cpp
@@ -260,55 +260,89 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP32]], 1
 // CHECK1-NEXT:    [[TMP33:%.*]] = zext i32 [[ADD]] to i64
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP33]])
-// CHECK1-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l28.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP29]], i32 0)
-// CHECK1-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK1-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP35]], align 4
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP40]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP41]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 [[TMP29]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK1-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l28(i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], [100 x i32]* @a) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP36:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[TMP44:%.*]] = load i32, i32* [[N_ADDR]], align 4
 // CHECK1-NEXT:    [[CONV7:%.*]] = bitcast i64* [[N_CASTED6]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP36]], i32* [[CONV7]], align 4
-// CHECK1-NEXT:    [[TMP37:%.*]] = load i64, i64* [[N_CASTED6]], align 8
-// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP37]], i64* [[TMP39]], align 8
-// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP37]], i64* [[TMP41]], align 8
-// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP42]], align 8
-// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [100 x i32]**
-// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP44]], align 8
-// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to [100 x i32]**
-// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP46]], align 8
-// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP47]], align 8
-// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP50:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    store i32 [[TMP50]], i32* [[DOTCAPTURE_EXPR_12]], align 4
-// CHECK1-NEXT:    [[TMP51:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_12]], align 4
-// CHECK1-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[TMP51]], 0
+// CHECK1-NEXT:    store i32 [[TMP44]], i32* [[CONV7]], align 4
+// CHECK1-NEXT:    [[TMP45:%.*]] = load i64, i64* [[N_CASTED6]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP45]], i64* [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP45]], i64* [[TMP49]], align 8
+// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP50]], align 8
+// CHECK1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to [100 x i32]**
+// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP52]], align 8
+// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [100 x i32]**
+// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP54]], align 8
+// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP55]], align 8
+// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP58:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    store i32 [[TMP58]], i32* [[DOTCAPTURE_EXPR_12]], align 4
+// CHECK1-NEXT:    [[TMP59:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_12]], align 4
+// CHECK1-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[TMP59]], 0
 // CHECK1-NEXT:    [[DIV15:%.*]] = sdiv i32 [[SUB14]], 1
 // CHECK1-NEXT:    [[SUB16:%.*]] = sub nsw i32 [[DIV15]], 1
 // CHECK1-NEXT:    store i32 [[SUB16]], i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK1-NEXT:    [[TMP52:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK1-NEXT:    [[ADD17:%.*]] = add nsw i32 [[TMP52]], 1
-// CHECK1-NEXT:    [[TMP53:%.*]] = zext i32 [[ADD17]] to i64
-// CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP53]])
-// CHECK1-NEXT:    [[TMP54:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35.region_id, i32 2, i8** [[TMP48]], i8** [[TMP49]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP55:%.*]] = icmp ne i32 [[TMP54]], 0
-// CHECK1-NEXT:    br i1 [[TMP55]], label [[OMP_OFFLOAD_FAILED18:%.*]], label [[OMP_OFFLOAD_CONT19:%.*]]
-// CHECK1:       omp_offload.failed18:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35(i64 [[TMP37]], [100 x i32]* @a) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT19]]
-// CHECK1:       omp_offload.cont19:
-// CHECK1-NEXT:    [[TMP56:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i64 0, i64 0), align 4
-// CHECK1-NEXT:    ret i32 [[TMP56]]
+// CHECK1-NEXT:    [[TMP60:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
+// CHECK1-NEXT:    [[ADD17:%.*]] = add nsw i32 [[TMP60]], 1
+// CHECK1-NEXT:    [[TMP61:%.*]] = zext i32 [[ADD17]] to i64
+// CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP61]])
+// CHECK1-NEXT:    [[KERNEL_ARGS18:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP62]], align 4
+// CHECK1-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP63]], align 4
+// CHECK1-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP56]], i8*** [[TMP64]], align 8
+// CHECK1-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP57]], i8*** [[TMP65]], align 8
+// CHECK1-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP66]], align 8
+// CHECK1-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP67]], align 8
+// CHECK1-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP68]], align 8
+// CHECK1-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP69]], align 8
+// CHECK1-NEXT:    [[TMP70:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS18]])
+// CHECK1-NEXT:    [[TMP71:%.*]] = icmp ne i32 [[TMP70]], 0
+// CHECK1-NEXT:    br i1 [[TMP71]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
+// CHECK1:       omp_offload.failed19:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35(i64 [[TMP45]], [100 x i32]* @a) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
+// CHECK1:       omp_offload.cont20:
+// CHECK1-NEXT:    [[TMP72:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i64 0, i64 0), align 4
+// CHECK1-NEXT:    ret i32 [[TMP72]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l28
@@ -831,54 +865,88 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP32]], 1
 // CHECK3-NEXT:    [[TMP33:%.*]] = zext i32 [[ADD]] to i64
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP33]])
-// CHECK3-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l28.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP29]], i32 0)
-// CHECK3-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK3-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 4
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 [[TMP29]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK3-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l28(i32 [[TMP2]], i32 [[TMP4]], i32 [[TMP6]], [100 x i32]* @a) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP36:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    store i32 [[TMP36]], i32* [[N_CASTED4]], align 4
-// CHECK3-NEXT:    [[TMP37:%.*]] = load i32, i32* [[N_CASTED4]], align 4
-// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP37]], i32* [[TMP39]], align 4
-// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP37]], i32* [[TMP41]], align 4
-// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP42]], align 4
-// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [100 x i32]**
-// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP44]], align 4
-// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to [100 x i32]**
-// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP46]], align 4
-// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP47]], align 4
-// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP50:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    store i32 [[TMP50]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK3-NEXT:    [[TMP51:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK3-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP51]], 0
+// CHECK3-NEXT:    [[TMP44:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    store i32 [[TMP44]], i32* [[N_CASTED4]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = load i32, i32* [[N_CASTED4]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP45]], i32* [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP45]], i32* [[TMP49]], align 4
+// CHECK3-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP50]], align 4
+// CHECK3-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to [100 x i32]**
+// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP52]], align 4
+// CHECK3-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [100 x i32]**
+// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP54]], align 4
+// CHECK3-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP55]], align 4
+// CHECK3-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP58:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    store i32 [[TMP58]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK3-NEXT:    [[TMP59:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK3-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP59]], 0
 // CHECK3-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK3-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK3-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK3-NEXT:    [[TMP52:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK3-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP52]], 1
-// CHECK3-NEXT:    [[TMP53:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP53]])
-// CHECK3-NEXT:    [[TMP54:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35.region_id, i32 2, i8** [[TMP48]], i8** [[TMP49]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP55:%.*]] = icmp ne i32 [[TMP54]], 0
-// CHECK3-NEXT:    br i1 [[TMP55]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK3:       omp_offload.failed15:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35(i32 [[TMP37]], [100 x i32]* @a) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK3:       omp_offload.cont16:
-// CHECK3-NEXT:    [[TMP56:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i32 0, i32 0), align 4
-// CHECK3-NEXT:    ret i32 [[TMP56]]
+// CHECK3-NEXT:    [[TMP60:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK3-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP60]], 1
+// CHECK3-NEXT:    [[TMP61:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP61]])
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP62]], align 4
+// CHECK3-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP63]], align 4
+// CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP56]], i8*** [[TMP64]], align 4
+// CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP57]], i8*** [[TMP65]], align 4
+// CHECK3-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP66]], align 4
+// CHECK3-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP67]], align 4
+// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP68]], align 4
+// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP71:%.*]] = icmp ne i32 [[TMP70]], 0
+// CHECK3-NEXT:    br i1 [[TMP71]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l35(i32 [[TMP45]], [100 x i32]* @a) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[TMP72:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i32 0, i32 0), align 4
+// CHECK3-NEXT:    ret i32 [[TMP72]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l28
@@ -1373,18 +1441,35 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK9-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK9-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 0
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP33]])
-// CHECK9-NEXT:    ret i32 [[TMP32]]
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP41]])
+// CHECK9-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73
@@ -1673,18 +1758,35 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK11-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK11-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK11-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK11-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i32 0
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP33]])
-// CHECK11-NEXT:    ret i32 [[TMP32]]
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP41]])
+// CHECK11-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l73
@@ -1930,17 +2032,34 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK17-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK17-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK17-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK17-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK17-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK17-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK17-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK17-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK17-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
 // CHECK17-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK17-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A2]], i64 0, i64 0
-// CHECK17-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK17-NEXT:    ret i32 [[TMP9]]
+// CHECK17-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK17-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109
@@ -2131,17 +2250,34 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK19-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK19-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK19-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK19-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK19-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK19-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK19-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK19-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK19-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
 // CHECK19-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK19-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A2]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK19-NEXT:    ret i32 [[TMP9]]
+// CHECK19-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK19-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l109
@@ -2371,20 +2507,37 @@ int main (int argc, char **argv) {
 // CHECK25-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK25-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK25-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK25-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK25-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK25-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK25-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK25-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK25-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK25-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK25-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK25-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK25-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK25-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK25-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK25-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK25-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK25-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK25-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK25-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK25-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK25-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK25-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK25-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK25-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK25-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK25:       omp_offload.failed:
 // CHECK25-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK25-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK25:       omp_offload.cont:
-// CHECK25-NEXT:    [[TMP32:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK25-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP32]])
+// CHECK25-NEXT:    [[TMP40:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK25-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP40]])
 // CHECK25-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK25-NEXT:    [[TMP33:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK25-NEXT:    call void @llvm.stackrestore(i8* [[TMP33]])
-// CHECK25-NEXT:    [[TMP34:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK25-NEXT:    ret i32 [[TMP34]]
+// CHECK25-NEXT:    [[TMP41:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK25-NEXT:    call void @llvm.stackrestore(i8* [[TMP41]])
+// CHECK25-NEXT:    [[TMP42:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK25-NEXT:    ret i32 [[TMP42]]
 //
 //
 // CHECK25-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162
@@ -2651,9 +2804,26 @@ int main (int argc, char **argv) {
 // CHECK25-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK25-NEXT:    [[TMP21:%.*]] = load i32, i32* [[TE]], align 4
 // CHECK25-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK25-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l151.region_id, i32 3, i8** [[TMP19]], i8** [[TMP20]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 [[TMP21]], i32 0)
-// CHECK25-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK25-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK25-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK25-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK25-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK25-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK25-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK25-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK25-NEXT:    store i8** [[TMP19]], i8*** [[TMP24]], align 8
+// CHECK25-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK25-NEXT:    store i8** [[TMP20]], i8*** [[TMP25]], align 8
+// CHECK25-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK25-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK25-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK25-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK25-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK25-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK25-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK25-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK25-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 [[TMP21]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l151.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK25-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK25-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK25:       omp_offload.failed:
 // CHECK25-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l151(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK25-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2901,20 +3071,37 @@ int main (int argc, char **argv) {
 // CHECK27-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK27-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK27-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK27-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK27-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK27-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK27-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK27-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK27-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK27-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK27-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK27-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK27-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK27-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK27-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK27-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK27-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK27-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK27-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK27-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK27-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK27-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK27-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK27-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK27-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK27-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK27:       omp_offload.failed:
 // CHECK27-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK27-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK27:       omp_offload.cont:
-// CHECK27-NEXT:    [[TMP32:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK27-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP32]])
+// CHECK27-NEXT:    [[TMP40:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK27-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP40]])
 // CHECK27-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK27-NEXT:    [[TMP33:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK27-NEXT:    call void @llvm.stackrestore(i8* [[TMP33]])
-// CHECK27-NEXT:    [[TMP34:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK27-NEXT:    ret i32 [[TMP34]]
+// CHECK27-NEXT:    [[TMP41:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK27-NEXT:    call void @llvm.stackrestore(i8* [[TMP41]])
+// CHECK27-NEXT:    [[TMP42:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK27-NEXT:    ret i32 [[TMP42]]
 //
 //
 // CHECK27-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l162
@@ -3173,9 +3360,26 @@ int main (int argc, char **argv) {
 // CHECK27-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK27-NEXT:    [[TMP21:%.*]] = load i32, i32* [[TE]], align 4
 // CHECK27-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK27-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l151.region_id, i32 3, i8** [[TMP19]], i8** [[TMP20]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 [[TMP21]], i32 0)
-// CHECK27-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK27-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK27-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK27-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK27-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK27-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK27-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK27-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK27-NEXT:    store i8** [[TMP19]], i8*** [[TMP24]], align 4
+// CHECK27-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK27-NEXT:    store i8** [[TMP20]], i8*** [[TMP25]], align 4
+// CHECK27-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK27-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK27-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK27-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK27-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK27-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK27-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK27-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK27-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 [[TMP21]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l151.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK27-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK27-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK27:       omp_offload.failed:
 // CHECK27-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l151(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK27-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_collapse_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_collapse_codegen.cpp
index bb3821d..a0a5e5f 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_collapse_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_collapse_codegen.cpp
@@ -133,9 +133,26 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 56088)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -143,8 +160,8 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i64 0, i64 0
 // CHECK1-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP9]]
+// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -352,9 +369,26 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 56088)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -362,8 +396,8 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP9]]
+// CHECK3-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -646,20 +680,37 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP46:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_4]], align 8
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP46]], 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK9-NEXT:    [[TMP47:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83.region_id, i32 5, i8** [[TMP39]], i8** [[TMP40]], i64* [[TMP41]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP48:%.*]] = icmp ne i32 [[TMP47]], 0
-// CHECK9-NEXT:    br i1 [[TMP48]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP47]], align 4
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP48]], align 4
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 8
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP40]], i8*** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP41]], i64** [[TMP51]], align 8
+// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP56:%.*]] = icmp ne i32 [[TMP55]], 0
+// CHECK9-NEXT:    br i1 [[TMP56]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83(i64 [[TMP7]], i64 [[TMP9]], i64 [[TMP1]], i64 [[TMP3]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP49:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP49]])
+// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP57]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP50:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP50]])
-// CHECK9-NEXT:    [[TMP51:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP51]]
+// CHECK9-NEXT:    [[TMP58:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP58]])
+// CHECK9-NEXT:    [[TMP59:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP59]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83
@@ -974,9 +1025,26 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 20)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l69.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l69.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l69([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1266,20 +1334,37 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP45:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_3]], align 8
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP45]], 1
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK11-NEXT:    [[TMP46:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83.region_id, i32 5, i8** [[TMP38]], i8** [[TMP39]], i64* [[TMP40]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP47:%.*]] = icmp ne i32 [[TMP46]], 0
-// CHECK11-NEXT:    br i1 [[TMP47]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP38]], i8*** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP40]], i64** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP55:%.*]] = icmp ne i32 [[TMP54]], 0
+// CHECK11-NEXT:    br i1 [[TMP55]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83(i32 [[TMP5]], i32 [[TMP7]], i32 [[TMP0]], i32 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP48:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP48]])
+// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP56]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP49]])
-// CHECK11-NEXT:    [[TMP50:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP50]]
+// CHECK11-NEXT:    [[TMP57:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP57]])
+// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP58]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83
@@ -1594,9 +1679,26 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 20)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l69.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l69.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l69([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_copyin_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_copyin_codegen.cpp
index 90dd355..d8d2ea7 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_copyin_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_copyin_codegen.cpp
@@ -126,9 +126,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l64.region_id, i32 2, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK1-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP14]], align 4
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP15]], align 4
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP12]], i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP13]], i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l64.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
+// CHECK1-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l64(i64 [[TMP1]], [2 x i32]* [[A]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -329,9 +346,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l34.region_id, i32 2, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK1-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP14]], align 4
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP15]], align 4
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP12]], i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP13]], i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l34.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
+// CHECK1-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l34(i64 [[TMP1]], [2 x i32]* [[A]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -544,9 +578,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l64.region_id, i32 2, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK3-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP12]], i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP13]], i8*** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP19]], align 4
+// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l64.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
+// CHECK3-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l64(i32 [[TMP1]], [2 x i32]* [[A]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -740,9 +791,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l34.region_id, i32 2, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK3-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP12]], i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP13]], i8*** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP19]], align 4
+// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l34.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
+// CHECK3-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l34(i32 [[TMP1]], [2 x i32]* [[A]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_dist_schedule_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_dist_schedule_codegen.cpp
index 701a72f..d3d5c5e 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_dist_schedule_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_dist_schedule_codegen.cpp
@@ -163,10 +163,10 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -181,55 +181,106 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK1:       omp_offload.failed16:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK1:       omp_offload.cont15:
-// CHECK1-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP27]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK1:       omp_offload.cont17:
+// CHECK1-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i64 0, i64 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -732,10 +783,10 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -750,55 +801,106 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK3:       omp_offload.failed8:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK3:       omp_offload.cont8:
-// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK3:       omp_offload.cont9:
+// CHECK3-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK3-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK3:       omp_offload.failed14:
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK3:       omp_offload.cont15:
-// CHECK3-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP27]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -1293,14 +1395,14 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[M_CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES24:%.*]] = alloca [4 x i64], align 8
-// CHECK9-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
+// CHECK9-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -1359,140 +1461,191 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK9-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK9-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK9-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
-// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK9-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK9-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK9-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK9-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK9-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK9-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK9-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP58]])
-// CHECK9-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l111.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK9-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK9:       omp_offload.failed16:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l111(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK9:       omp_offload.cont17:
-// CHECK9-NEXT:    [[TMP61:%.*]] = load i32, i32* [[M]], align 4
-// CHECK9-NEXT:    [[CONV18:%.*]] = bitcast i64* [[M_CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP61]], i32* [[CONV18]], align 4
-// CHECK9-NEXT:    [[TMP62:%.*]] = load i64, i64* [[M_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP63]], i32* [[CONV20]], align 4
-// CHECK9-NEXT:    [[TMP64:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK9-NEXT:    [[TMP65:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP66:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES24]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP66]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
-// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP62]], i64* [[TMP68]], align 8
-// CHECK9-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP62]], i64* [[TMP70]], align 8
-// CHECK9-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP71]], align 8
-// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP64]], i64* [[TMP73]], align 8
-// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP64]], i64* [[TMP75]], align 8
-// CHECK9-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP76]], align 8
-// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP78]], align 8
-// CHECK9-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP80]], align 8
-// CHECK9-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP81]], align 8
-// CHECK9-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 8
-// CHECK9-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP85]], align 8
-// CHECK9-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 3
-// CHECK9-NEXT:    store i64 [[TMP65]], i64* [[TMP86]], align 8
-// CHECK9-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK9-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK9-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP66]])
+// CHECK9-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK9-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK9-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK9-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK9-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l111.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK9-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK9-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK9:       omp_offload.failed17:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l111(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK9:       omp_offload.cont18:
+// CHECK9-NEXT:    [[TMP77:%.*]] = load i32, i32* [[M]], align 4
+// CHECK9-NEXT:    [[CONV19:%.*]] = bitcast i64* [[M_CASTED]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP77]], i32* [[CONV19]], align 4
+// CHECK9-NEXT:    [[TMP78:%.*]] = load i64, i64* [[M_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP79]], i32* [[CONV21]], align 4
+// CHECK9-NEXT:    [[TMP80:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK9-NEXT:    [[TMP81:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP82:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP82]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
+// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP78]], i64* [[TMP84]], align 8
+// CHECK9-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP78]], i64* [[TMP86]], align 8
+// CHECK9-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
 // CHECK9-NEXT:    store i8* null, i8** [[TMP87]], align 8
-// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP91:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP91]], i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK9-NEXT:    [[TMP92:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK9-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[TMP92]], 0
-// CHECK9-NEXT:    [[DIV29:%.*]] = sdiv i32 [[SUB28]], 1
-// CHECK9-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[DIV29]], 1
-// CHECK9-NEXT:    store i32 [[SUB30]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK9-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK9-NEXT:    [[ADD31:%.*]] = add nsw i32 [[TMP93]], 1
-// CHECK9-NEXT:    [[TMP94:%.*]] = zext i32 [[ADD31]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP94]])
-// CHECK9-NEXT:    [[TMP95:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116.region_id, i32 4, i8** [[TMP88]], i8** [[TMP89]], i64* [[TMP90]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP96:%.*]] = icmp ne i32 [[TMP95]], 0
-// CHECK9-NEXT:    br i1 [[TMP96]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
-// CHECK9:       omp_offload.failed32:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116(i64 [[TMP62]], i64 [[TMP64]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
-// CHECK9:       omp_offload.cont33:
-// CHECK9-NEXT:    [[TMP97:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP97]])
+// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP80]], i64* [[TMP89]], align 8
+// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP80]], i64* [[TMP91]], align 8
+// CHECK9-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP92]], align 8
+// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP94]], align 8
+// CHECK9-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP96]], align 8
+// CHECK9-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP97]], align 8
+// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 8
+// CHECK9-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP101]], align 8
+// CHECK9-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 3
+// CHECK9-NEXT:    store i64 [[TMP81]], i64* [[TMP102]], align 8
+// CHECK9-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP103]], align 8
+// CHECK9-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP107:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP107]], i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK9-NEXT:    [[TMP108:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK9-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP108]], 0
+// CHECK9-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
+// CHECK9-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
+// CHECK9-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK9-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK9-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP109]], 1
+// CHECK9-NEXT:    [[TMP110:%.*]] = zext i32 [[ADD32]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP110]])
+// CHECK9-NEXT:    [[KERNEL_ARGS33:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP111]], align 4
+// CHECK9-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP112]], align 4
+// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP104]], i8*** [[TMP113]], align 8
+// CHECK9-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP106]], i64** [[TMP115]], align 8
+// CHECK9-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP116]], align 8
+// CHECK9-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP117]], align 8
+// CHECK9-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK9-NEXT:    [[TMP119:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]])
+// CHECK9-NEXT:    [[TMP120:%.*]] = icmp ne i32 [[TMP119]], 0
+// CHECK9-NEXT:    br i1 [[TMP120]], label [[OMP_OFFLOAD_FAILED34:%.*]], label [[OMP_OFFLOAD_CONT35:%.*]]
+// CHECK9:       omp_offload.failed34:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116(i64 [[TMP78]], i64 [[TMP80]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT35]]
+// CHECK9:       omp_offload.cont35:
+// CHECK9-NEXT:    [[TMP121:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP121]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP98:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP98]])
-// CHECK9-NEXT:    [[TMP99:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP99]]
+// CHECK9-NEXT:    [[TMP122:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP122]])
+// CHECK9-NEXT:    [[TMP123:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP123]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106
@@ -2193,10 +2346,10 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[M_CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -2210,61 +2363,112 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK9-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK9:       omp_offload.failed5:
+// CHECK9-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK9-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK9-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK9:       omp_offload.failed6:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK9:       omp_offload.cont6:
-// CHECK9-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK9:       omp_offload.cont7:
+// CHECK9-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[M_CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP18]], i32* [[CONV]], align 4
-// CHECK9-NEXT:    [[TMP19:%.*]] = load i64, i64* [[M_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP19]], i64* [[TMP21]], align 8
-// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP19]], i64* [[TMP23]], align 8
-// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP24]], align 8
-// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP26]], align 8
-// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP28]], align 8
-// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP29]], align 8
-// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 [[TMP34]], i32* [[CONV]], align 4
+// CHECK9-NEXT:    [[TMP35:%.*]] = load i64, i64* [[M_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP35]], i64* [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP35]], i64* [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP44]], align 8
+// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l94.region_id, i32 2, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK9-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK9:       omp_offload.failed11:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l94(i64 [[TMP19]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK9:       omp_offload.cont12:
+// CHECK9-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 2, i32* [[TMP49]], align 4
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP46]], i8*** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 8
+// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK9-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK9-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK9:       omp_offload.failed13:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l94(i64 [[TMP35]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK9:       omp_offload.cont14:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -2795,14 +2999,14 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[M_CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED17:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS18:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES21:%.*]] = alloca [4 x i64], align 4
-// CHECK11-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_23:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
+// CHECK11-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -2860,139 +3064,190 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK11-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK11-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK11-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK11-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK11-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
-// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK11-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK11-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK11-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK11-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK11-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK11-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP59]])
-// CHECK11-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l111.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK11-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK11:       omp_offload.failed15:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l111(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK11:       omp_offload.cont16:
-// CHECK11-NEXT:    [[TMP62:%.*]] = load i32, i32* [[M]], align 4
-// CHECK11-NEXT:    store i32 [[TMP62]], i32* [[M_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP63:%.*]] = load i32, i32* [[M_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[N_CASTED17]], align 4
-// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[N_CASTED17]], align 4
-// CHECK11-NEXT:    [[TMP66:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP67:%.*]] = sext i32 [[TMP66]] to i64
-// CHECK11-NEXT:    [[TMP68:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES21]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP68]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
-// CHECK11-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP63]], i32* [[TMP70]], align 4
-// CHECK11-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP63]], i32* [[TMP72]], align 4
-// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP73]], align 4
-// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP65]], i32* [[TMP75]], align 4
-// CHECK11-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP65]], i32* [[TMP77]], align 4
-// CHECK11-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP78]], align 4
-// CHECK11-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP80]], align 4
-// CHECK11-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP82]], align 4
-// CHECK11-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP83]], align 4
-// CHECK11-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP85]], align 4
-// CHECK11-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP87]], align 4
-// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 3
-// CHECK11-NEXT:    store i64 [[TMP67]], i64* [[TMP88]], align 4
-// CHECK11-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK11-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP67]])
+// CHECK11-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK11-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK11-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK11-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK11-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l111.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK11-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK11-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK11:       omp_offload.failed16:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l111(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK11:       omp_offload.cont17:
+// CHECK11-NEXT:    [[TMP78:%.*]] = load i32, i32* [[M]], align 4
+// CHECK11-NEXT:    store i32 [[TMP78]], i32* [[M_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP79:%.*]] = load i32, i32* [[M_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP80]], i32* [[N_CASTED18]], align 4
+// CHECK11-NEXT:    [[TMP81:%.*]] = load i32, i32* [[N_CASTED18]], align 4
+// CHECK11-NEXT:    [[TMP82:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP83:%.*]] = sext i32 [[TMP82]] to i64
+// CHECK11-NEXT:    [[TMP84:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP84]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
+// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP79]], i32* [[TMP86]], align 4
+// CHECK11-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP88:%.*]] = bitcast i8** [[TMP87]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP79]], i32* [[TMP88]], align 4
+// CHECK11-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
 // CHECK11-NEXT:    store i8* null, i8** [[TMP89]], align 4
-// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP93:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP93]], i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK11-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK11-NEXT:    [[SUB25:%.*]] = sub nsw i32 [[TMP94]], 0
-// CHECK11-NEXT:    [[DIV26:%.*]] = sdiv i32 [[SUB25]], 1
-// CHECK11-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[DIV26]], 1
-// CHECK11-NEXT:    store i32 [[SUB27]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[ADD28:%.*]] = add nsw i32 [[TMP95]], 1
-// CHECK11-NEXT:    [[TMP96:%.*]] = zext i32 [[ADD28]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP96]])
-// CHECK11-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116.region_id, i32 4, i8** [[TMP90]], i8** [[TMP91]], i64* [[TMP92]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
-// CHECK11-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
-// CHECK11:       omp_offload.failed29:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116(i32 [[TMP63]], i32 [[TMP65]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
-// CHECK11:       omp_offload.cont30:
-// CHECK11-NEXT:    [[TMP99:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP99]])
+// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP81]], i32* [[TMP91]], align 4
+// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP81]], i32* [[TMP93]], align 4
+// CHECK11-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP94]], align 4
+// CHECK11-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP96]], align 4
+// CHECK11-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP98]], align 4
+// CHECK11-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP99]], align 4
+// CHECK11-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP101]], align 4
+// CHECK11-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP103]], align 4
+// CHECK11-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 3
+// CHECK11-NEXT:    store i64 [[TMP83]], i64* [[TMP104]], align 4
+// CHECK11-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP105]], align 4
+// CHECK11-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP109:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP109]], i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK11-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK11-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP110]], 0
+// CHECK11-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
+// CHECK11-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
+// CHECK11-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP111]], 1
+// CHECK11-NEXT:    [[TMP112:%.*]] = zext i32 [[ADD29]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP112]])
+// CHECK11-NEXT:    [[KERNEL_ARGS30:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP113]], align 4
+// CHECK11-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP114]], align 4
+// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 4
+// CHECK11-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK11-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP108]], i64** [[TMP117]], align 4
+// CHECK11-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP118]], align 4
+// CHECK11-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP119]], align 4
+// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK11-NEXT:    [[TMP121:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]])
+// CHECK11-NEXT:    [[TMP122:%.*]] = icmp ne i32 [[TMP121]], 0
+// CHECK11-NEXT:    br i1 [[TMP122]], label [[OMP_OFFLOAD_FAILED31:%.*]], label [[OMP_OFFLOAD_CONT32:%.*]]
+// CHECK11:       omp_offload.failed31:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116(i32 [[TMP79]], i32 [[TMP81]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT32]]
+// CHECK11:       omp_offload.cont32:
+// CHECK11-NEXT:    [[TMP123:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP123]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP100:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP100]])
-// CHECK11-NEXT:    [[TMP101:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP101]]
+// CHECK11-NEXT:    [[TMP124:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP124]])
+// CHECK11-NEXT:    [[TMP125:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP125]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106
@@ -3670,10 +3925,10 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 4
 // CHECK11-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[M_CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -3687,60 +3942,111 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK11-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK11-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK11:       omp_offload.failed5:
+// CHECK11-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK11-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK11-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK11:       omp_offload.failed6:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK11:       omp_offload.cont6:
-// CHECK11-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK11-NEXT:    store i32 [[TMP18]], i32* [[M_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP19:%.*]] = load i32, i32* [[M_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP19]], i32* [[TMP21]], align 4
-// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP19]], i32* [[TMP23]], align 4
-// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP24]], align 4
-// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP26]], align 4
-// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP28]], align 4
-// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP29]], align 4
-// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK11:       omp_offload.cont7:
+// CHECK11-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK11-NEXT:    store i32 [[TMP34]], i32* [[M_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = load i32, i32* [[M_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP35]], i32* [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP35]], i32* [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP44]], align 4
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP45]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l94.region_id, i32 2, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK11-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK11:       omp_offload.failed11:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l94(i32 [[TMP19]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK11:       omp_offload.cont12:
+// CHECK11-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 2, i32* [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP46]], i8*** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP55]], align 4
+// CHECK11-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK11-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK11-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK11:       omp_offload.failed13:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l94(i32 [[TMP35]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK11:       omp_offload.cont14:
 // CHECK11-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_firstprivate_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_firstprivate_codegen.cpp
index d2064a4..42847b6 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_firstprivate_codegen.cpp
@@ -351,9 +351,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK1-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK1-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99(i64 [[TMP1]], [2 x i32]* @vec, [2 x %struct.S]* @s_arr, %struct.S* @var, i64 [[TMP3]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -764,27 +781,44 @@ int main() {
 // CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK1-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK1-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP32]]
+// CHECK1-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN2StC2Ev
@@ -1413,9 +1447,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK3-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK3-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l99(i32 [[TMP1]], [2 x i32]* @vec, [2 x %struct.S]* @s_arr, %struct.S* @var, i32 [[TMP3]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1809,27 +1860,44 @@ int main() {
 // CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK3-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK3-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP32]]
+// CHECK3-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN2StC2Ev

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_if_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_if_codegen.cpp
index 8e56321..d10494d 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_if_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_if_codegen.cpp
@@ -119,21 +119,55 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l48() #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l52() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret void
 //
 //
@@ -428,55 +462,106 @@ int main() {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i8, align 1
-// CHECK1-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l81() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l89.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l89.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l89() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
+// CHECK1-NEXT:    [[TMP20:%.*]] = load i32, i32* @Arg, align 4
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP13:%.*]] = load i32, i32* @Arg, align 4
-// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP13]], 0
+// CHECK1-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP29:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP29]], 0
 // CHECK1-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP14:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP14]] to i1
-// CHECK1-NEXT:    [[TMP15:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP30:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL5:%.*]] = trunc i8 [[TMP30]] to i1
+// CHECK1-NEXT:    [[TMP31:%.*]] = select i1 [[TOBOOL5]], i32 0, i32 1
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l97.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP15]])
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK1:       omp_offload.failed6:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l97(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK1:       omp_offload.cont7:
-// CHECK1-NEXT:    [[TMP18:%.*]] = load i32, i32* @Arg, align 4
-// CHECK1-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP18]])
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP31]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l97.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK1-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l97(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[TMP42:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP42]])
 // CHECK1-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -942,53 +1027,104 @@ int main() {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i8, align 1
-// CHECK1-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
+// CHECK1-NEXT:    [[TMP20:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP13:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
-// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP13]], 0
+// CHECK1-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP29:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP29]], 0
 // CHECK1-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP14:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP14]] to i1
-// CHECK1-NEXT:    [[TMP15:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP30:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL5:%.*]] = trunc i8 [[TMP30]] to i1
+// CHECK1-NEXT:    [[TMP31:%.*]] = select i1 [[TOBOOL5]], i32 0, i32 1
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l72.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP15]])
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK1:       omp_offload.failed6:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l72(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK1:       omp_offload.cont7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP31]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l72.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK1-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l72(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
 // CHECK1-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_lastprivate_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_lastprivate_codegen.cpp
index 102a2d5..a5fca51 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_lastprivate_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_lastprivate_codegen.cpp
@@ -797,9 +797,26 @@ int main() {
 // CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK9-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK9-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -807,18 +824,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done3:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP39]]
+// CHECK9-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1258,27 +1275,44 @@ int main() {
 // CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK9-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK9-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP32]]
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1794,9 +1828,26 @@ int main() {
 // CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK11-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK11-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l106(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1804,18 +1855,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP39]]
+// CHECK11-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -2246,27 +2297,44 @@ int main() {
 // CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK11-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK11-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP32]]
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_num_threads_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_num_threads_codegen.cpp
index e930e2e..af75839 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_num_threads_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_num_threads_codegen.cpp
@@ -95,68 +95,102 @@ int main() {
 // CHECK1:       invoke.cont:
 // CHECK1-NEXT:    store i8 [[CALL]], i8* [[A]], align 1
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l50.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 2)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 2, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l50.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l50() #[[ATTR6:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       lpad:
-// CHECK1-NEXT:    [[TMP2:%.*]] = landingpad { i8*, i32 }
+// CHECK1-NEXT:    [[TMP10:%.*]] = landingpad { i8*, i32 }
 // CHECK1-NEXT:    cleanup
-// CHECK1-NEXT:    [[TMP3:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 0
-// CHECK1-NEXT:    store i8* [[TMP3]], i8** [[EXN_SLOT]], align 8
-// CHECK1-NEXT:    [[TMP4:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 1
-// CHECK1-NEXT:    store i32 [[TMP4]], i32* [[EHSELECTOR_SLOT]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 0
+// CHECK1-NEXT:    store i8* [[TMP11]], i8** [[EXN_SLOT]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 1
+// CHECK1-NEXT:    store i32 [[TMP12]], i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK1-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
 // CHECK1-NEXT:    br label [[EH_RESUME:%.*]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i8, i8* [[A]], align 1
+// CHECK1-NEXT:    [[TMP13:%.*]] = load i8, i8* [[A]], align 1
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i8*
-// CHECK1-NEXT:    store i8 [[TMP5]], i8* [[CONV]], align 1
-// CHECK1-NEXT:    [[TMP6:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP8]], align 8
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP11]], align 8
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP14:%.*]] = load i8, i8* [[A]], align 1
-// CHECK1-NEXT:    store i8 [[TMP14]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP15:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP16:%.*]] = zext i8 [[TMP15]] to i32
+// CHECK1-NEXT:    store i8 [[TMP13]], i8* [[CONV]], align 1
+// CHECK1-NEXT:    [[TMP14:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP22:%.*]] = load i8, i8* [[A]], align 1
+// CHECK1-NEXT:    store i8 [[TMP22]], i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TMP23:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TMP24:%.*]] = zext i8 [[TMP23]] to i32
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55.region_id, i32 1, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP16]])
-// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
-// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55(i64 [[TMP6]]) #[[ATTR6]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
-// CHECK1-NEXT:    [[TMP19:%.*]] = load i8, i8* [[A]], align 1
-// CHECK1-NEXT:    [[CONV4:%.*]] = sext i8 [[TMP19]] to i32
-// CHECK1-NEXT:    [[CALL6:%.*]] = invoke noundef signext i32 @_Z5tmainIcLi5EEiv()
-// CHECK1-NEXT:    to label [[INVOKE_CONT5:%.*]] unwind label [[LPAD]]
-// CHECK1:       invoke.cont5:
-// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV4]], [[CALL6]]
-// CHECK1-NEXT:    [[CALL8:%.*]] = invoke noundef signext i32 @_Z5tmainI1SLi1EEiv()
-// CHECK1-NEXT:    to label [[INVOKE_CONT7:%.*]] unwind label [[LPAD]]
-// CHECK1:       invoke.cont7:
-// CHECK1-NEXT:    [[ADD9:%.*]] = add nsw i32 [[ADD]], [[CALL8]]
-// CHECK1-NEXT:    store i32 [[ADD9]], i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP24]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
+// CHECK1-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55(i64 [[TMP14]]) #[[ATTR6]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
+// CHECK1-NEXT:    [[TMP35:%.*]] = load i8, i8* [[A]], align 1
+// CHECK1-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP35]] to i32
+// CHECK1-NEXT:    [[CALL7:%.*]] = invoke noundef signext i32 @_Z5tmainIcLi5EEiv()
+// CHECK1-NEXT:    to label [[INVOKE_CONT6:%.*]] unwind label [[LPAD]]
+// CHECK1:       invoke.cont6:
+// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV5]], [[CALL7]]
+// CHECK1-NEXT:    [[CALL9:%.*]] = invoke noundef signext i32 @_Z5tmainI1SLi1EEiv()
+// CHECK1-NEXT:    to label [[INVOKE_CONT8:%.*]] unwind label [[LPAD]]
+// CHECK1:       invoke.cont8:
+// CHECK1-NEXT:    [[ADD10:%.*]] = add nsw i32 [[ADD]], [[CALL9]]
+// CHECK1-NEXT:    store i32 [[ADD10]], i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
-// CHECK1-NEXT:    [[TMP20:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP20]]
+// CHECK1-NEXT:    [[TMP36:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP36]]
 // CHECK1:       eh.resume:
 // CHECK1-NEXT:    [[EXN:%.*]] = load i8*, i8** [[EXN_SLOT]], align 8
 // CHECK1-NEXT:    [[SEL:%.*]] = load i32, i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK1-NEXT:    [[LPAD_VAL:%.*]] = insertvalue { i8*, i32 } undef, i8* [[EXN]], 0
-// CHECK1-NEXT:    [[LPAD_VAL10:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
-// CHECK1-NEXT:    resume { i8*, i32 } [[LPAD_VAL10]]
+// CHECK1-NEXT:    [[LPAD_VAL11:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
+// CHECK1-NEXT:    resume { i8*, i32 } [[LPAD_VAL11]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN1SC1El
@@ -507,21 +541,55 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l36.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 5)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 5, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l36() #[[ATTR6]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l40.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 23)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 23, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l40() #[[ATTR6]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret i32 0
 //
 //
@@ -533,9 +601,26 @@ int main() {
 // CHECK1-NEXT:    [[REF_TMP:%.*]] = alloca [[STRUCT_S:%.*]], align 8
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l36.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l36() #[[ATTR6]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -546,22 +631,39 @@ int main() {
 // CHECK1-NEXT:    [[CALL:%.*]] = call noundef signext i8 @_ZN1ScvcEv(%struct.S* noundef nonnull align 8 dereferenceable(24) [[REF_TMP]])
 // CHECK1-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[REF_TMP]]) #[[ATTR6]]
 // CHECK1-NEXT:    store i8 [[CALL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP2:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP3:%.*]] = zext i8 [[TMP2]] to i32
+// CHECK1-NEXT:    [[TMP10:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TMP11:%.*]] = zext i8 [[TMP10]] to i32
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP4:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l40.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 [[TMP3]])
-// CHECK1-NEXT:    [[TMP5:%.*]] = icmp ne i32 [[TMP4]], 0
-// CHECK1-NEXT:    br i1 [[TMP5]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP13]], align 4
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP11]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
+// CHECK1-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l40() #[[ATTR6]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret i32 0
 // CHECK1:       terminate.lpad:
-// CHECK1-NEXT:    [[TMP6:%.*]] = landingpad { i8*, i32 }
+// CHECK1-NEXT:    [[TMP22:%.*]] = landingpad { i8*, i32 }
 // CHECK1-NEXT:    catch i8* null
-// CHECK1-NEXT:    [[TMP7:%.*]] = extractvalue { i8*, i32 } [[TMP6]], 0
-// CHECK1-NEXT:    call void @__clang_call_terminate(i8* [[TMP7]]) #[[ATTR9]]
+// CHECK1-NEXT:    [[TMP23:%.*]] = extractvalue { i8*, i32 } [[TMP22]], 0
+// CHECK1-NEXT:    call void @__clang_call_terminate(i8* [[TMP23]]) #[[ATTR9]]
 // CHECK1-NEXT:    unreachable
 //
 //
@@ -1239,68 +1341,102 @@ int main() {
 // CHECK5:       invoke.cont:
 // CHECK5-NEXT:    store i8 [[CALL]], i8* [[A]], align 1
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l50.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 2)
-// CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK5-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK5-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK5-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK5-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK5-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 2, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l50.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK5-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l50() #[[ATTR6:[0-9]+]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       lpad:
-// CHECK5-NEXT:    [[TMP2:%.*]] = landingpad { i8*, i32 }
+// CHECK5-NEXT:    [[TMP10:%.*]] = landingpad { i8*, i32 }
 // CHECK5-NEXT:    cleanup
-// CHECK5-NEXT:    [[TMP3:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 0
-// CHECK5-NEXT:    store i8* [[TMP3]], i8** [[EXN_SLOT]], align 8
-// CHECK5-NEXT:    [[TMP4:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 1
-// CHECK5-NEXT:    store i32 [[TMP4]], i32* [[EHSELECTOR_SLOT]], align 4
+// CHECK5-NEXT:    [[TMP11:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 0
+// CHECK5-NEXT:    store i8* [[TMP11]], i8** [[EXN_SLOT]], align 8
+// CHECK5-NEXT:    [[TMP12:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 1
+// CHECK5-NEXT:    store i32 [[TMP12]], i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK5-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
 // CHECK5-NEXT:    br label [[EH_RESUME:%.*]]
 // CHECK5:       omp_offload.cont:
-// CHECK5-NEXT:    [[TMP5:%.*]] = load i8, i8* [[A]], align 1
+// CHECK5-NEXT:    [[TMP13:%.*]] = load i8, i8* [[A]], align 1
 // CHECK5-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i8*
-// CHECK5-NEXT:    store i8 [[TMP5]], i8* [[CONV]], align 1
-// CHECK5-NEXT:    [[TMP6:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP6]], i64* [[TMP8]], align 8
-// CHECK5-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP6]], i64* [[TMP10]], align 8
-// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP11]], align 8
-// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP14:%.*]] = load i8, i8* [[A]], align 1
-// CHECK5-NEXT:    store i8 [[TMP14]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK5-NEXT:    [[TMP15:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK5-NEXT:    [[TMP16:%.*]] = zext i8 [[TMP15]] to i32
+// CHECK5-NEXT:    store i8 [[TMP13]], i8* [[CONV]], align 1
+// CHECK5-NEXT:    [[TMP14:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP14]], i64* [[TMP16]], align 8
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP14]], i64* [[TMP18]], align 8
+// CHECK5-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP19]], align 8
+// CHECK5-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP22:%.*]] = load i8, i8* [[A]], align 1
+// CHECK5-NEXT:    store i8 [[TMP22]], i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK5-NEXT:    [[TMP23:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK5-NEXT:    [[TMP24:%.*]] = zext i8 [[TMP23]] to i32
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55.region_id, i32 1, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP16]])
-// CHECK5-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
-// CHECK5-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK5:       omp_offload.failed2:
-// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55(i64 [[TMP6]]) #[[ATTR6]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK5:       omp_offload.cont3:
-// CHECK5-NEXT:    [[TMP19:%.*]] = load i8, i8* [[A]], align 1
-// CHECK5-NEXT:    [[CONV4:%.*]] = sext i8 [[TMP19]] to i32
-// CHECK5-NEXT:    [[CALL6:%.*]] = invoke noundef signext i32 @_Z5tmainIcLi5EEiv()
-// CHECK5-NEXT:    to label [[INVOKE_CONT5:%.*]] unwind label [[LPAD]]
-// CHECK5:       invoke.cont5:
-// CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV4]], [[CALL6]]
-// CHECK5-NEXT:    [[CALL8:%.*]] = invoke noundef signext i32 @_Z5tmainI1SLi1EEiv()
-// CHECK5-NEXT:    to label [[INVOKE_CONT7:%.*]] unwind label [[LPAD]]
-// CHECK5:       invoke.cont7:
-// CHECK5-NEXT:    [[ADD9:%.*]] = add nsw i32 [[ADD]], [[CALL8]]
-// CHECK5-NEXT:    store i32 [[ADD9]], i32* [[RETVAL]], align 4
+// CHECK5-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP20]], i8*** [[TMP27]], align 8
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP21]], i8*** [[TMP28]], align 8
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP30]], align 8
+// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP32]], align 8
+// CHECK5-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP24]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK5-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
+// CHECK5-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK5:       omp_offload.failed3:
+// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55(i64 [[TMP14]]) #[[ATTR6]]
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK5:       omp_offload.cont4:
+// CHECK5-NEXT:    [[TMP35:%.*]] = load i8, i8* [[A]], align 1
+// CHECK5-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP35]] to i32
+// CHECK5-NEXT:    [[CALL7:%.*]] = invoke noundef signext i32 @_Z5tmainIcLi5EEiv()
+// CHECK5-NEXT:    to label [[INVOKE_CONT6:%.*]] unwind label [[LPAD]]
+// CHECK5:       invoke.cont6:
+// CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV5]], [[CALL7]]
+// CHECK5-NEXT:    [[CALL9:%.*]] = invoke noundef signext i32 @_Z5tmainI1SLi1EEiv()
+// CHECK5-NEXT:    to label [[INVOKE_CONT8:%.*]] unwind label [[LPAD]]
+// CHECK5:       invoke.cont8:
+// CHECK5-NEXT:    [[ADD10:%.*]] = add nsw i32 [[ADD]], [[CALL9]]
+// CHECK5-NEXT:    store i32 [[ADD10]], i32* [[RETVAL]], align 4
 // CHECK5-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
-// CHECK5-NEXT:    [[TMP20:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP20]]
+// CHECK5-NEXT:    [[TMP36:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP36]]
 // CHECK5:       eh.resume:
 // CHECK5-NEXT:    [[EXN:%.*]] = load i8*, i8** [[EXN_SLOT]], align 8
 // CHECK5-NEXT:    [[SEL:%.*]] = load i32, i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK5-NEXT:    [[LPAD_VAL:%.*]] = insertvalue { i8*, i32 } undef, i8* [[EXN]], 0
-// CHECK5-NEXT:    [[LPAD_VAL10:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
-// CHECK5-NEXT:    resume { i8*, i32 } [[LPAD_VAL10]]
+// CHECK5-NEXT:    [[LPAD_VAL11:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
+// CHECK5-NEXT:    resume { i8*, i32 } [[LPAD_VAL11]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@_ZN1SC1El
@@ -1651,21 +1787,55 @@ int main() {
 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l36.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 5)
-// CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK5-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK5-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK5-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK5-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK5-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 5, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK5-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l36() #[[ATTR6]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l40.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 23)
-// CHECK5-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK5-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK5:       omp_offload.failed2:
+// CHECK5-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK5-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK5-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 23, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK5-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK5-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK5:       omp_offload.failed3:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l40() #[[ATTR6]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK5:       omp_offload.cont3:
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK5:       omp_offload.cont4:
 // CHECK5-NEXT:    ret i32 0
 //
 //
@@ -1677,9 +1847,26 @@ int main() {
 // CHECK5-NEXT:    [[REF_TMP:%.*]] = alloca [[STRUCT_S:%.*]], align 8
 // CHECK5-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l36.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK5-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK5-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK5-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK5-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK5-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK5-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l36() #[[ATTR6]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1690,22 +1877,39 @@ int main() {
 // CHECK5-NEXT:    [[CALL:%.*]] = call noundef signext i8 @_ZN1ScvcEv(%struct.S* noundef nonnull align 8 dereferenceable(24) [[REF_TMP]])
 // CHECK5-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[REF_TMP]]) #[[ATTR6]]
 // CHECK5-NEXT:    store i8 [[CALL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK5-NEXT:    [[TMP2:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK5-NEXT:    [[TMP3:%.*]] = zext i8 [[TMP2]] to i32
+// CHECK5-NEXT:    [[TMP10:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK5-NEXT:    [[TMP11:%.*]] = zext i8 [[TMP10]] to i32
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP4:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l40.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 [[TMP3]])
-// CHECK5-NEXT:    [[TMP5:%.*]] = icmp ne i32 [[TMP4]], 0
-// CHECK5-NEXT:    br i1 [[TMP5]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK5:       omp_offload.failed2:
+// CHECK5-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP13]], align 4
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK5-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP16]], align 8
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP17]], align 8
+// CHECK5-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP18]], align 8
+// CHECK5-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP19]], align 8
+// CHECK5-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP11]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK5-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
+// CHECK5-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK5:       omp_offload.failed3:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l40() #[[ATTR6]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK5:       omp_offload.cont3:
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK5:       omp_offload.cont4:
 // CHECK5-NEXT:    ret i32 0
 // CHECK5:       terminate.lpad:
-// CHECK5-NEXT:    [[TMP6:%.*]] = landingpad { i8*, i32 }
+// CHECK5-NEXT:    [[TMP22:%.*]] = landingpad { i8*, i32 }
 // CHECK5-NEXT:    catch i8* null
-// CHECK5-NEXT:    [[TMP7:%.*]] = extractvalue { i8*, i32 } [[TMP6]], 0
-// CHECK5-NEXT:    call void @__clang_call_terminate(i8* [[TMP7]]) #[[ATTR9]]
+// CHECK5-NEXT:    [[TMP23:%.*]] = extractvalue { i8*, i32 } [[TMP22]], 0
+// CHECK5-NEXT:    call void @__clang_call_terminate(i8* [[TMP23]]) #[[ATTR9]]
 // CHECK5-NEXT:    unreachable
 //
 //

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_private_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_private_codegen.cpp
index 73fd95d..1ddecc8 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_private_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_private_codegen.cpp
@@ -262,9 +262,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -509,27 +526,44 @@ int main() {
 // CHECK1-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 8
 // CHECK1-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 8
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK1-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP4]]
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev
@@ -955,9 +989,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1196,27 +1247,44 @@ int main() {
 // CHECK3-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 4
 // CHECK3-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK3-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP4]]
+// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_proc_bind_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_proc_bind_codegen.cpp
index ef6a946..5cd958d 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_proc_bind_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_proc_bind_codegen.cpp
@@ -58,21 +58,55 @@ int main() {
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l36.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l36() #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l39.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l39.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l39() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK1-NEXT:    ret i32 [[CALL]]
 //
@@ -358,9 +392,26 @@ int main() {
 // CHECK1-NEXT:  entry:
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l29.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l29.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l29() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_reduction_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_reduction_codegen.cpp
index 89195f9..d37086b 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_reduction_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_reduction_codegen.cpp
@@ -118,9 +118,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68(i64 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -403,9 +420,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP10:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[TMP10]], 0
-// CHECK1-NEXT:    br i1 [[TMP11]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP9]], i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i64 [[TMP2]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -690,9 +724,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK3-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l68(i32 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -969,9 +1020,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP10:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[TMP10]], 0
-// CHECK3-NEXT:    br i1 [[TMP11]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP9]], i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK3-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i32 [[TMP2]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_schedule_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_schedule_codegen.cpp
index 1a22efc..f284d2f 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_schedule_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_schedule_codegen.cpp
@@ -228,18 +228,18 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -254,93 +254,178 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l45.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l45.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK1:       omp_offload.failed16:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l45(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK1:       omp_offload.cont15:
-// CHECK1-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 8
-// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 8
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP31]], align 8
-// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK1:       omp_offload.cont17:
+// CHECK1-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 8
+// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 8
+// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP55]], align 8
+// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l51.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK1-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK1:       omp_offload.failed21:
+// CHECK1-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK1-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK1-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 8
+// CHECK1-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 8
+// CHECK1-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 8
+// CHECK1-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 8
+// CHECK1-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP64]], align 8
+// CHECK1-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP65]], align 8
+// CHECK1-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l51.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK1-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK1-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK1:       omp_offload.failed24:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l51(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK1:       omp_offload.cont22:
-// CHECK1-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 8
-// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 8
-// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK1:       omp_offload.cont25:
+// CHECK1-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 8
+// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP72]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l57.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK1-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK1:       omp_offload.failed28:
+// CHECK1-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK1-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 8
+// CHECK1-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 8
+// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 8
+// CHECK1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 8
+// CHECK1-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP81]], align 8
+// CHECK1-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP82]], align 8
+// CHECK1-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK1-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK1-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK1:       omp_offload.failed32:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l57(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK1:       omp_offload.cont29:
-// CHECK1-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP45]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK1:       omp_offload.cont33:
+// CHECK1-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i64 0, i64 0
+// CHECK1-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35
@@ -1142,18 +1227,18 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -1168,93 +1253,178 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK3:       omp_offload.failed8:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK3:       omp_offload.cont8:
-// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK3:       omp_offload.cont9:
+// CHECK3-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l45.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK3-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK3:       omp_offload.failed14:
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l45.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l45(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK3:       omp_offload.cont15:
-// CHECK3-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 4
-// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 4
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP31]], align 4
-// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 4
+// CHECK3-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 4
+// CHECK3-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP55]], align 4
+// CHECK3-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l51.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK3-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK3:       omp_offload.failed21:
+// CHECK3-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK3-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK3-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 4
+// CHECK3-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 4
+// CHECK3-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 4
+// CHECK3-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 4
+// CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP64]], align 4
+// CHECK3-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP65]], align 4
+// CHECK3-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l51.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK3-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK3-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK3:       omp_offload.failed24:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l51(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK3:       omp_offload.cont22:
-// CHECK3-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 4
-// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 4
-// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP40]], align 4
-// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK3:       omp_offload.cont25:
+// CHECK3-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 4
+// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP72]], align 4
+// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l57.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK3-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK3:       omp_offload.failed28:
+// CHECK3-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK3-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK3-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 4
+// CHECK3-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 4
+// CHECK3-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 4
+// CHECK3-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 4
+// CHECK3-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP81]], align 4
+// CHECK3-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP82]], align 4
+// CHECK3-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK3-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK3-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK3:       omp_offload.failed32:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l57(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK3:       omp_offload.cont29:
-// CHECK3-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP45]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK3:       omp_offload.cont33:
+// CHECK3-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35
@@ -2029,18 +2199,18 @@ int main (int argc, char **argv) {
 // CHECK5-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK5-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 8
-// CHECK5-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 8
+// CHECK5-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK5-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK5-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -2055,93 +2225,178 @@ int main (int argc, char **argv) {
 // CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK5-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK5-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK5-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK5-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK5-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
 // CHECK5-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK5-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK5-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK5-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK5-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK5:       omp_offload.failed7:
+// CHECK5-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK5-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK5-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK5-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK5:       omp_offload.failed8:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK5:       omp_offload.cont8:
-// CHECK5-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK5-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK5-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK5-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK5:       omp_offload.cont9:
+// CHECK5-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK5-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK5-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l45.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK5-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK5:       omp_offload.failed14:
+// CHECK5-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK5-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK5-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK5-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK5-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK5-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK5-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK5-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK5-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l45.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK5-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK5-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK5:       omp_offload.failed16:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l45(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK5:       omp_offload.cont15:
-// CHECK5-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 8
-// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 8
-// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP31]], align 8
-// CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK5:       omp_offload.cont17:
+// CHECK5-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 8
+// CHECK5-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 8
+// CHECK5-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP55]], align 8
+// CHECK5-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l51.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK5-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK5:       omp_offload.failed21:
+// CHECK5-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK5-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK5-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 8
+// CHECK5-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 8
+// CHECK5-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 8
+// CHECK5-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 8
+// CHECK5-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP64]], align 8
+// CHECK5-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP65]], align 8
+// CHECK5-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l51.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK5-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK5-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK5:       omp_offload.failed24:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l51(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK5:       omp_offload.cont22:
-// CHECK5-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 8
-// CHECK5-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 8
-// CHECK5-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK5-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK5:       omp_offload.cont25:
+// CHECK5-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 8
+// CHECK5-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 8
+// CHECK5-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP72]], align 8
+// CHECK5-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l57.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK5-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK5:       omp_offload.failed28:
+// CHECK5-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK5-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK5-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 8
+// CHECK5-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 8
+// CHECK5-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 8
+// CHECK5-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 8
+// CHECK5-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP81]], align 8
+// CHECK5-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP82]], align 8
+// CHECK5-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK5-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK5-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK5:       omp_offload.failed32:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l57(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK5:       omp_offload.cont29:
-// CHECK5-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i64 0, i64 0
-// CHECK5-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP45]]
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK5:       omp_offload.cont33:
+// CHECK5-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i64 0, i64 0
+// CHECK5-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35
@@ -2943,18 +3198,18 @@ int main (int argc, char **argv) {
 // CHECK7-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK7-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 4
-// CHECK7-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 4
+// CHECK7-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK7-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK7-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK7-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -2969,93 +3224,178 @@ int main (int argc, char **argv) {
 // CHECK7-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK7-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK7-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK7-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK7-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK7-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK7-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK7-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK7-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK7-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK7-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK7-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK7-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK7-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK7-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK7:       omp_offload.failed:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK7:       omp_offload.cont:
 // CHECK7-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK7-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK7-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK7-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK7-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK7-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK7-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK7-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK7-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK7-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK7:       omp_offload.failed7:
+// CHECK7-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK7-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK7-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK7-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK7-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK7-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK7-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK7-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK7-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK7:       omp_offload.failed8:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l40(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK7:       omp_offload.cont8:
-// CHECK7-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK7-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK7-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK7-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK7-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK7:       omp_offload.cont9:
+// CHECK7-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK7-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK7-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK7-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK7-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l45.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK7-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK7:       omp_offload.failed14:
+// CHECK7-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK7-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK7-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK7-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK7-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK7-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK7-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK7-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK7-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l45.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK7-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK7-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK7:       omp_offload.failed16:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l45(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK7:       omp_offload.cont15:
-// CHECK7-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 4
-// CHECK7-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK7-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 4
-// CHECK7-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP31]], align 4
-// CHECK7-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK7:       omp_offload.cont17:
+// CHECK7-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 4
+// CHECK7-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK7-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 4
+// CHECK7-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP55]], align 4
+// CHECK7-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l51.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK7-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK7:       omp_offload.failed21:
+// CHECK7-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK7-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK7-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 4
+// CHECK7-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 4
+// CHECK7-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 4
+// CHECK7-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 4
+// CHECK7-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP64]], align 4
+// CHECK7-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP65]], align 4
+// CHECK7-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l51.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK7-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK7-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK7:       omp_offload.failed24:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l51(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK7:       omp_offload.cont22:
-// CHECK7-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 4
-// CHECK7-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK7-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 4
-// CHECK7-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
-// CHECK7-NEXT:    store i8* null, i8** [[TMP40]], align 4
-// CHECK7-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK7:       omp_offload.cont25:
+// CHECK7-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK7-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 4
+// CHECK7-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK7-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 4
+// CHECK7-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i32 0, i32 0
+// CHECK7-NEXT:    store i8* null, i8** [[TMP72]], align 4
+// CHECK7-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK7-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK7-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l57.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK7-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK7-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK7:       omp_offload.failed28:
+// CHECK7-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK7-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK7-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK7-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK7-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK7-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK7-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 4
+// CHECK7-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK7-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 4
+// CHECK7-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 4
+// CHECK7-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK7-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 4
+// CHECK7-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP81]], align 4
+// CHECK7-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK7-NEXT:    store i8** null, i8*** [[TMP82]], align 4
+// CHECK7-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK7-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK7-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK7:       omp_offload.failed32:
 // CHECK7-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l57(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK7:       omp_offload.cont29:
-// CHECK7-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i32 0, i32 0
-// CHECK7-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK7-NEXT:    ret i32 [[TMP45]]
+// CHECK7-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK7:       omp_offload.cont33:
+// CHECK7-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK7-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i32 0, i32 0
+// CHECK7-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK7-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK7-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l35
@@ -3837,31 +4177,31 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[M_CASTED:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES24:%.*]] = alloca [4 x i64], align 8
-// CHECK13-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
+// CHECK13-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[N_CASTED34:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS36:%.*]] = alloca [3 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS37:%.*]] = alloca [3 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS38:%.*]] = alloca [3 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES39:%.*]] = alloca [3 x i64], align 8
-// CHECK13-NEXT:    [[_TMP40:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_41:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_42:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[M_CASTED49:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[N_CASTED51:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS53:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS54:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS55:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES56:%.*]] = alloca [4 x i64], align 8
-// CHECK13-NEXT:    [[_TMP57:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_58:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_59:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[N_CASTED36:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS38:%.*]] = alloca [3 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS39:%.*]] = alloca [3 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS40:%.*]] = alloca [3 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES41:%.*]] = alloca [3 x i64], align 8
+// CHECK13-NEXT:    [[_TMP42:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_43:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_44:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[M_CASTED52:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[N_CASTED54:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS56:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS57:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS58:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES59:%.*]] = alloca [4 x i64], align 8
+// CHECK13-NEXT:    [[_TMP60:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_61:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_62:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK13-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK13-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -3920,260 +4260,345 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK13-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK13-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK13-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK13-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK13-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK13-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK13-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK13-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK13-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK13-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK13-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK13-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       omp_offload.cont:
-// CHECK13-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK13-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK13-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK13-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
-// CHECK13-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK13-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK13-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK13-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK13-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK13-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK13-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK13-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK13-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK13-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK13-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK13-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK13-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK13-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK13-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK13-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
+// CHECK13-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK13-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK13-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK13-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK13-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK13-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK13-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK13-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK13-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK13-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK13-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK13-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK13-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK13-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK13-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK13-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK13-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK13-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP58]])
-// CHECK13-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK13-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK13:       omp_offload.failed16:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK13:       omp_offload.cont17:
-// CHECK13-NEXT:    [[TMP61:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    [[CONV18:%.*]] = bitcast i64* [[M_CASTED]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP61]], i32* [[CONV18]], align 4
-// CHECK13-NEXT:    [[TMP62:%.*]] = load i64, i64* [[M_CASTED]], align 8
-// CHECK13-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP63]], i32* [[CONV20]], align 4
-// CHECK13-NEXT:    [[TMP64:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK13-NEXT:    [[TMP65:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP66:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES24]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP66]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
-// CHECK13-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP62]], i64* [[TMP68]], align 8
-// CHECK13-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP62]], i64* [[TMP70]], align 8
-// CHECK13-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP71]], align 8
-// CHECK13-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP64]], i64* [[TMP73]], align 8
-// CHECK13-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP64]], i64* [[TMP75]], align 8
-// CHECK13-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP76]], align 8
-// CHECK13-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP78]], align 8
-// CHECK13-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP80]], align 8
-// CHECK13-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP81]], align 8
-// CHECK13-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
-// CHECK13-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 8
-// CHECK13-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
-// CHECK13-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP85]], align 8
-// CHECK13-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 3
-// CHECK13-NEXT:    store i64 [[TMP65]], i64* [[TMP86]], align 8
-// CHECK13-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK13-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK13-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK13-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP66]])
+// CHECK13-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK13-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK13-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK13-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK13-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK13-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK13-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK13-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK13-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK13-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK13-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK13:       omp_offload.failed17:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK13:       omp_offload.cont18:
+// CHECK13-NEXT:    [[TMP77:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    [[CONV19:%.*]] = bitcast i64* [[M_CASTED]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP77]], i32* [[CONV19]], align 4
+// CHECK13-NEXT:    [[TMP78:%.*]] = load i64, i64* [[M_CASTED]], align 8
+// CHECK13-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP79]], i32* [[CONV21]], align 4
+// CHECK13-NEXT:    [[TMP80:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK13-NEXT:    [[TMP81:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP82:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP82]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
+// CHECK13-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP78]], i64* [[TMP84]], align 8
+// CHECK13-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP78]], i64* [[TMP86]], align 8
+// CHECK13-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
 // CHECK13-NEXT:    store i8* null, i8** [[TMP87]], align 8
-// CHECK13-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP91:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP91]], i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK13-NEXT:    [[TMP92:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK13-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[TMP92]], 0
-// CHECK13-NEXT:    [[DIV29:%.*]] = sdiv i32 [[SUB28]], 1
-// CHECK13-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[DIV29]], 1
-// CHECK13-NEXT:    store i32 [[SUB30]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK13-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK13-NEXT:    [[ADD31:%.*]] = add nsw i32 [[TMP93]], 1
-// CHECK13-NEXT:    [[TMP94:%.*]] = zext i32 [[ADD31]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP94]])
-// CHECK13-NEXT:    [[TMP95:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158.region_id, i32 4, i8** [[TMP88]], i8** [[TMP89]], i64* [[TMP90]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP96:%.*]] = icmp ne i32 [[TMP95]], 0
-// CHECK13-NEXT:    br i1 [[TMP96]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
-// CHECK13:       omp_offload.failed32:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158(i64 [[TMP62]], i64 [[TMP64]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
-// CHECK13:       omp_offload.cont33:
-// CHECK13-NEXT:    [[TMP97:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    [[CONV35:%.*]] = bitcast i64* [[N_CASTED34]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP97]], i32* [[CONV35]], align 4
-// CHECK13-NEXT:    [[TMP98:%.*]] = load i64, i64* [[N_CASTED34]], align 8
-// CHECK13-NEXT:    [[TMP99:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP100:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES39]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP100]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
-// CHECK13-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP98]], i64* [[TMP102]], align 8
-// CHECK13-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP98]], i64* [[TMP104]], align 8
-// CHECK13-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP105]], align 8
-// CHECK13-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP107]], align 8
-// CHECK13-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP109]], align 8
-// CHECK13-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP110]], align 8
-// CHECK13-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP112]], align 8
-// CHECK13-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP114]], align 8
-// CHECK13-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES39]], i32 0, i32 2
-// CHECK13-NEXT:    store i64 [[TMP99]], i64* [[TMP115]], align 8
-// CHECK13-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP116]], align 8
-// CHECK13-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES39]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP120:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP120]], i32* [[DOTCAPTURE_EXPR_41]], align 4
-// CHECK13-NEXT:    [[TMP121:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_41]], align 4
-// CHECK13-NEXT:    [[SUB43:%.*]] = sub nsw i32 [[TMP121]], 0
-// CHECK13-NEXT:    [[DIV44:%.*]] = sdiv i32 [[SUB43]], 1
-// CHECK13-NEXT:    [[SUB45:%.*]] = sub nsw i32 [[DIV44]], 1
-// CHECK13-NEXT:    store i32 [[SUB45]], i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK13-NEXT:    [[TMP122:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK13-NEXT:    [[ADD46:%.*]] = add nsw i32 [[TMP122]], 1
-// CHECK13-NEXT:    [[TMP123:%.*]] = zext i32 [[ADD46]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP123]])
-// CHECK13-NEXT:    [[TMP124:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163.region_id, i32 3, i8** [[TMP117]], i8** [[TMP118]], i64* [[TMP119]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP125:%.*]] = icmp ne i32 [[TMP124]], 0
-// CHECK13-NEXT:    br i1 [[TMP125]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
-// CHECK13:       omp_offload.failed47:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163(i64 [[TMP98]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
-// CHECK13:       omp_offload.cont48:
-// CHECK13-NEXT:    [[TMP126:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    [[CONV50:%.*]] = bitcast i64* [[M_CASTED49]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP126]], i32* [[CONV50]], align 4
-// CHECK13-NEXT:    [[TMP127:%.*]] = load i64, i64* [[M_CASTED49]], align 8
-// CHECK13-NEXT:    [[TMP128:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    [[CONV52:%.*]] = bitcast i64* [[N_CASTED51]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP128]], i32* [[CONV52]], align 4
-// CHECK13-NEXT:    [[TMP129:%.*]] = load i64, i64* [[N_CASTED51]], align 8
-// CHECK13-NEXT:    [[TMP130:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP131:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES56]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP131]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
-// CHECK13-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP80]], i64* [[TMP89]], align 8
+// CHECK13-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP80]], i64* [[TMP91]], align 8
+// CHECK13-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP92]], align 8
+// CHECK13-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP94]], align 8
+// CHECK13-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP96]], align 8
+// CHECK13-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP97]], align 8
+// CHECK13-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 8
+// CHECK13-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP101]], align 8
+// CHECK13-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 3
+// CHECK13-NEXT:    store i64 [[TMP81]], i64* [[TMP102]], align 8
+// CHECK13-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK13-NEXT:    store i8* null, i8** [[TMP103]], align 8
+// CHECK13-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP107:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP107]], i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK13-NEXT:    [[TMP108:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK13-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP108]], 0
+// CHECK13-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
+// CHECK13-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
+// CHECK13-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK13-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK13-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP109]], 1
+// CHECK13-NEXT:    [[TMP110:%.*]] = zext i32 [[ADD32]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP110]])
+// CHECK13-NEXT:    [[KERNEL_ARGS33:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP111]], align 4
+// CHECK13-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 4, i32* [[TMP112]], align 4
+// CHECK13-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP104]], i8*** [[TMP113]], align 8
+// CHECK13-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK13-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP106]], i64** [[TMP115]], align 8
+// CHECK13-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP116]], align 8
+// CHECK13-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP117]], align 8
+// CHECK13-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK13-NEXT:    [[TMP119:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]])
+// CHECK13-NEXT:    [[TMP120:%.*]] = icmp ne i32 [[TMP119]], 0
+// CHECK13-NEXT:    br i1 [[TMP120]], label [[OMP_OFFLOAD_FAILED34:%.*]], label [[OMP_OFFLOAD_CONT35:%.*]]
+// CHECK13:       omp_offload.failed34:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158(i64 [[TMP78]], i64 [[TMP80]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT35]]
+// CHECK13:       omp_offload.cont35:
+// CHECK13-NEXT:    [[TMP121:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[CONV37:%.*]] = bitcast i64* [[N_CASTED36]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP121]], i32* [[CONV37]], align 4
+// CHECK13-NEXT:    [[TMP122:%.*]] = load i64, i64* [[N_CASTED36]], align 8
+// CHECK13-NEXT:    [[TMP123:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP124:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES41]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP124]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
+// CHECK13-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP122]], i64* [[TMP126]], align 8
+// CHECK13-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP122]], i64* [[TMP128]], align 8
+// CHECK13-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS40]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP129]], align 8
+// CHECK13-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP131]], align 8
+// CHECK13-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 1
 // CHECK13-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP127]], i64* [[TMP133]], align 8
-// CHECK13-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP127]], i64* [[TMP135]], align 8
-// CHECK13-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP136]], align 8
-// CHECK13-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP129]], i64* [[TMP138]], align 8
-// CHECK13-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP129]], i64* [[TMP140]], align 8
-// CHECK13-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP141]], align 8
-// CHECK13-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP143]], align 8
-// CHECK13-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP145:%.*]] = bitcast i8** [[TMP144]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP145]], align 8
-// CHECK13-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP146]], align 8
-// CHECK13-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 3
-// CHECK13-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP148]], align 8
-// CHECK13-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 3
-// CHECK13-NEXT:    [[TMP150:%.*]] = bitcast i8** [[TMP149]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP150]], align 8
-// CHECK13-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES56]], i32 0, i32 3
-// CHECK13-NEXT:    store i64 [[TMP130]], i64* [[TMP151]], align 8
-// CHECK13-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 3
-// CHECK13-NEXT:    store i8* null, i8** [[TMP152]], align 8
-// CHECK13-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES56]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP156:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP156]], i32* [[DOTCAPTURE_EXPR_58]], align 4
-// CHECK13-NEXT:    [[TMP157:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_58]], align 4
-// CHECK13-NEXT:    [[SUB60:%.*]] = sub nsw i32 [[TMP157]], 0
-// CHECK13-NEXT:    [[DIV61:%.*]] = sdiv i32 [[SUB60]], 1
-// CHECK13-NEXT:    [[SUB62:%.*]] = sub nsw i32 [[DIV61]], 1
-// CHECK13-NEXT:    store i32 [[SUB62]], i32* [[DOTCAPTURE_EXPR_59]], align 4
-// CHECK13-NEXT:    [[TMP158:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_59]], align 4
-// CHECK13-NEXT:    [[ADD63:%.*]] = add nsw i32 [[TMP158]], 1
-// CHECK13-NEXT:    [[TMP159:%.*]] = zext i32 [[ADD63]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP159]])
-// CHECK13-NEXT:    [[TMP160:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168.region_id, i32 4, i8** [[TMP153]], i8** [[TMP154]], i64* [[TMP155]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP161:%.*]] = icmp ne i32 [[TMP160]], 0
-// CHECK13-NEXT:    br i1 [[TMP161]], label [[OMP_OFFLOAD_FAILED64:%.*]], label [[OMP_OFFLOAD_CONT65:%.*]]
-// CHECK13:       omp_offload.failed64:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168(i64 [[TMP127]], i64 [[TMP129]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT65]]
-// CHECK13:       omp_offload.cont65:
-// CHECK13-NEXT:    [[TMP162:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK13-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP162]])
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP133]], align 8
+// CHECK13-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS40]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP134]], align 8
+// CHECK13-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP136]], align 8
+// CHECK13-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP138]], align 8
+// CHECK13-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES41]], i32 0, i32 2
+// CHECK13-NEXT:    store i64 [[TMP123]], i64* [[TMP139]], align 8
+// CHECK13-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS40]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP140]], align 8
+// CHECK13-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES41]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP144:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP144]], i32* [[DOTCAPTURE_EXPR_43]], align 4
+// CHECK13-NEXT:    [[TMP145:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_43]], align 4
+// CHECK13-NEXT:    [[SUB45:%.*]] = sub nsw i32 [[TMP145]], 0
+// CHECK13-NEXT:    [[DIV46:%.*]] = sdiv i32 [[SUB45]], 1
+// CHECK13-NEXT:    [[SUB47:%.*]] = sub nsw i32 [[DIV46]], 1
+// CHECK13-NEXT:    store i32 [[SUB47]], i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK13-NEXT:    [[TMP146:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK13-NEXT:    [[ADD48:%.*]] = add nsw i32 [[TMP146]], 1
+// CHECK13-NEXT:    [[TMP147:%.*]] = zext i32 [[ADD48]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP147]])
+// CHECK13-NEXT:    [[KERNEL_ARGS49:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP148]], align 4
+// CHECK13-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 3, i32* [[TMP149]], align 4
+// CHECK13-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP141]], i8*** [[TMP150]], align 8
+// CHECK13-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP142]], i8*** [[TMP151]], align 8
+// CHECK13-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP143]], i64** [[TMP152]], align 8
+// CHECK13-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP153]], align 8
+// CHECK13-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP154]], align 8
+// CHECK13-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP155]], align 8
+// CHECK13-NEXT:    [[TMP156:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]])
+// CHECK13-NEXT:    [[TMP157:%.*]] = icmp ne i32 [[TMP156]], 0
+// CHECK13-NEXT:    br i1 [[TMP157]], label [[OMP_OFFLOAD_FAILED50:%.*]], label [[OMP_OFFLOAD_CONT51:%.*]]
+// CHECK13:       omp_offload.failed50:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163(i64 [[TMP122]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT51]]
+// CHECK13:       omp_offload.cont51:
+// CHECK13-NEXT:    [[TMP158:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    [[CONV53:%.*]] = bitcast i64* [[M_CASTED52]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP158]], i32* [[CONV53]], align 4
+// CHECK13-NEXT:    [[TMP159:%.*]] = load i64, i64* [[M_CASTED52]], align 8
+// CHECK13-NEXT:    [[TMP160:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[CONV55:%.*]] = bitcast i64* [[N_CASTED54]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP160]], i32* [[CONV55]], align 4
+// CHECK13-NEXT:    [[TMP161:%.*]] = load i64, i64* [[N_CASTED54]], align 8
+// CHECK13-NEXT:    [[TMP162:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP163:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES59]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP163]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
+// CHECK13-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP165:%.*]] = bitcast i8** [[TMP164]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP159]], i64* [[TMP165]], align 8
+// CHECK13-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP159]], i64* [[TMP167]], align 8
+// CHECK13-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP168]], align 8
+// CHECK13-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP170:%.*]] = bitcast i8** [[TMP169]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP161]], i64* [[TMP170]], align 8
+// CHECK13-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP172:%.*]] = bitcast i8** [[TMP171]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP161]], i64* [[TMP172]], align 8
+// CHECK13-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP173]], align 8
+// CHECK13-NEXT:    [[TMP174:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP175:%.*]] = bitcast i8** [[TMP174]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP175]], align 8
+// CHECK13-NEXT:    [[TMP176:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP177:%.*]] = bitcast i8** [[TMP176]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP177]], align 8
+// CHECK13-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP178]], align 8
+// CHECK13-NEXT:    [[TMP179:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP180:%.*]] = bitcast i8** [[TMP179]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP180]], align 8
+// CHECK13-NEXT:    [[TMP181:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP182:%.*]] = bitcast i8** [[TMP181]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP182]], align 8
+// CHECK13-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES59]], i32 0, i32 3
+// CHECK13-NEXT:    store i64 [[TMP162]], i64* [[TMP183]], align 8
+// CHECK13-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 3
+// CHECK13-NEXT:    store i8* null, i8** [[TMP184]], align 8
+// CHECK13-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES59]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP188:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP188]], i32* [[DOTCAPTURE_EXPR_61]], align 4
+// CHECK13-NEXT:    [[TMP189:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_61]], align 4
+// CHECK13-NEXT:    [[SUB63:%.*]] = sub nsw i32 [[TMP189]], 0
+// CHECK13-NEXT:    [[DIV64:%.*]] = sdiv i32 [[SUB63]], 1
+// CHECK13-NEXT:    [[SUB65:%.*]] = sub nsw i32 [[DIV64]], 1
+// CHECK13-NEXT:    store i32 [[SUB65]], i32* [[DOTCAPTURE_EXPR_62]], align 4
+// CHECK13-NEXT:    [[TMP190:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_62]], align 4
+// CHECK13-NEXT:    [[ADD66:%.*]] = add nsw i32 [[TMP190]], 1
+// CHECK13-NEXT:    [[TMP191:%.*]] = zext i32 [[ADD66]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP191]])
+// CHECK13-NEXT:    [[KERNEL_ARGS67:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP192]], align 4
+// CHECK13-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 4, i32* [[TMP193]], align 4
+// CHECK13-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP185]], i8*** [[TMP194]], align 8
+// CHECK13-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP186]], i8*** [[TMP195]], align 8
+// CHECK13-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP187]], i64** [[TMP196]], align 8
+// CHECK13-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP197]], align 8
+// CHECK13-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP198]], align 8
+// CHECK13-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP199]], align 8
+// CHECK13-NEXT:    [[TMP200:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]])
+// CHECK13-NEXT:    [[TMP201:%.*]] = icmp ne i32 [[TMP200]], 0
+// CHECK13-NEXT:    br i1 [[TMP201]], label [[OMP_OFFLOAD_FAILED68:%.*]], label [[OMP_OFFLOAD_CONT69:%.*]]
+// CHECK13:       omp_offload.failed68:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168(i64 [[TMP159]], i64 [[TMP161]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT69]]
+// CHECK13:       omp_offload.cont69:
+// CHECK13-NEXT:    [[TMP202:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK13-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP202]])
 // CHECK13-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK13-NEXT:    [[TMP163:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK13-NEXT:    call void @llvm.stackrestore(i8* [[TMP163]])
-// CHECK13-NEXT:    [[TMP164:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK13-NEXT:    ret i32 [[TMP164]]
+// CHECK13-NEXT:    [[TMP203:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK13-NEXT:    call void @llvm.stackrestore(i8* [[TMP203]])
+// CHECK13-NEXT:    [[TMP204:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK13-NEXT:    ret i32 [[TMP204]]
 //
 //
 // CHECK13-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148
@@ -5313,19 +5738,19 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 8
 // CHECK13-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[M_CASTED:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 8
-// CHECK13-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[M_CASTED19:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 8
+// CHECK13-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[M_CASTED22:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK13-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK13-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -5339,109 +5764,194 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK13-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK13-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK13-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK13-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK13-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK13-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK13-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK13-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK13-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       omp_offload.cont:
-// CHECK13-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK13-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK13-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK13-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK13-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l121.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK13-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK13:       omp_offload.failed5:
+// CHECK13-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK13-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK13-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK13-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK13-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK13-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK13-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK13-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK13-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l121.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK13-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK13-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK13:       omp_offload.failed6:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l121([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK13:       omp_offload.cont6:
-// CHECK13-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK13:       omp_offload.cont7:
+// CHECK13-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
 // CHECK13-NEXT:    [[CONV:%.*]] = bitcast i64* [[M_CASTED]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP18]], i32* [[CONV]], align 4
-// CHECK13-NEXT:    [[TMP19:%.*]] = load i64, i64* [[M_CASTED]], align 8
-// CHECK13-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP19]], i64* [[TMP21]], align 8
-// CHECK13-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP19]], i64* [[TMP23]], align 8
-// CHECK13-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP24]], align 8
-// CHECK13-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP26]], align 8
-// CHECK13-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP28]], align 8
-// CHECK13-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP29]], align 8
-// CHECK13-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 [[TMP34]], i32* [[CONV]], align 4
+// CHECK13-NEXT:    [[TMP35:%.*]] = load i64, i64* [[M_CASTED]], align 8
+// CHECK13-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP35]], i64* [[TMP37]], align 8
+// CHECK13-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP35]], i64* [[TMP39]], align 8
+// CHECK13-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP40]], align 8
+// CHECK13-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP42]], align 8
+// CHECK13-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP44]], align 8
+// CHECK13-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP45]], align 8
+// CHECK13-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126.region_id, i32 2, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK13-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK13:       omp_offload.failed11:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126(i64 [[TMP19]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK13:       omp_offload.cont12:
-// CHECK13-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 8
-// CHECK13-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 8
-// CHECK13-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP38]], align 8
-// CHECK13-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK13-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK13-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 2, i32* [[TMP49]], align 4
+// CHECK13-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP46]], i8*** [[TMP50]], align 8
+// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 8
+// CHECK13-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK13-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK13-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK13-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK13-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK13-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK13:       omp_offload.failed13:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126(i64 [[TMP35]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK13:       omp_offload.cont14:
+// CHECK13-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP59]], align 8
+// CHECK13-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP61]], align 8
+// CHECK13-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP62]], align 8
+// CHECK13-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP41:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l131.region_id, i32 1, i8** [[TMP39]], i8** [[TMP40]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP42:%.*]] = icmp ne i32 [[TMP41]], 0
-// CHECK13-NEXT:    br i1 [[TMP42]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK13:       omp_offload.failed17:
+// CHECK13-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP65]], align 4
+// CHECK13-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK13-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP63]], i8*** [[TMP67]], align 8
+// CHECK13-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 8
+// CHECK13-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP69]], align 8
+// CHECK13-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP70]], align 8
+// CHECK13-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP71]], align 8
+// CHECK13-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP72]], align 8
+// CHECK13-NEXT:    [[TMP73:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l131.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK13-NEXT:    [[TMP74:%.*]] = icmp ne i32 [[TMP73]], 0
+// CHECK13-NEXT:    br i1 [[TMP74]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK13:       omp_offload.failed20:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l131([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK13:       omp_offload.cont18:
-// CHECK13-NEXT:    [[TMP43:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    [[CONV20:%.*]] = bitcast i64* [[M_CASTED19]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP43]], i32* [[CONV20]], align 4
-// CHECK13-NEXT:    [[TMP44:%.*]] = load i64, i64* [[M_CASTED19]], align 8
-// CHECK13-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP44]], i64* [[TMP46]], align 8
-// CHECK13-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP44]], i64* [[TMP48]], align 8
-// CHECK13-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP49]], align 8
-// CHECK13-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP51]], align 8
-// CHECK13-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP53]], align 8
-// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP54]], align 8
-// CHECK13-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK13:       omp_offload.cont21:
+// CHECK13-NEXT:    [[TMP75:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    [[CONV23:%.*]] = bitcast i64* [[M_CASTED22]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP75]], i32* [[CONV23]], align 4
+// CHECK13-NEXT:    [[TMP76:%.*]] = load i64, i64* [[M_CASTED22]], align 8
+// CHECK13-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP76]], i64* [[TMP78]], align 8
+// CHECK13-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP76]], i64* [[TMP80]], align 8
+// CHECK13-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP81]], align 8
+// CHECK13-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP83]], align 8
+// CHECK13-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP85]], align 8
+// CHECK13-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP86]], align 8
+// CHECK13-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
-// CHECK13-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
-// CHECK13:       omp_offload.failed25:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136(i64 [[TMP44]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
-// CHECK13:       omp_offload.cont26:
+// CHECK13-NEXT:    [[KERNEL_ARGS28:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP89]], align 4
+// CHECK13-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 2, i32* [[TMP90]], align 4
+// CHECK13-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP87]], i8*** [[TMP91]], align 8
+// CHECK13-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP88]], i8*** [[TMP92]], align 8
+// CHECK13-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP93]], align 8
+// CHECK13-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP94]], align 8
+// CHECK13-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP95]], align 8
+// CHECK13-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP96]], align 8
+// CHECK13-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]])
+// CHECK13-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
+// CHECK13-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
+// CHECK13:       omp_offload.failed29:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136(i64 [[TMP76]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
+// CHECK13:       omp_offload.cont30:
 // CHECK13-NEXT:    ret i32 0
 //
 //
@@ -6292,31 +6802,31 @@ int main (int argc, char **argv) {
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[M_CASTED:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[N_CASTED17:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS18:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES21:%.*]] = alloca [4 x i64], align 4
-// CHECK15-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_23:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
+// CHECK15-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[N_CASTED31:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS32:%.*]] = alloca [3 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS33:%.*]] = alloca [3 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS34:%.*]] = alloca [3 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES35:%.*]] = alloca [3 x i64], align 4
-// CHECK15-NEXT:    [[_TMP36:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_37:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_38:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[M_CASTED45:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[N_CASTED46:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS47:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS48:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS49:%.*]] = alloca [4 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES50:%.*]] = alloca [4 x i64], align 4
-// CHECK15-NEXT:    [[_TMP51:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_52:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_53:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[N_CASTED33:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS34:%.*]] = alloca [3 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS35:%.*]] = alloca [3 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS36:%.*]] = alloca [3 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES37:%.*]] = alloca [3 x i64], align 4
+// CHECK15-NEXT:    [[_TMP38:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_39:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_40:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[M_CASTED48:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[N_CASTED49:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS50:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS51:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS52:%.*]] = alloca [4 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_SIZES53:%.*]] = alloca [4 x i64], align 4
+// CHECK15-NEXT:    [[_TMP54:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_55:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTCAPTURE_EXPR_56:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK15-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK15-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -6374,258 +6884,343 @@ int main (int argc, char **argv) {
 // CHECK15-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK15-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK15-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK15-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK15-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK15-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK15-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK15-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK15-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK15-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK15-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK15-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK15-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK15-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK15-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK15-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK15:       omp_offload.failed:
 // CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK15:       omp_offload.cont:
-// CHECK15-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK15-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK15-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK15-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK15-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
-// CHECK15-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK15-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK15-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK15-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK15-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK15-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK15-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK15-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK15-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK15-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK15-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK15-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK15-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK15-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK15-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK15-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK15-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK15-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK15-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK15-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
+// CHECK15-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK15-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK15-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK15-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK15-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK15-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK15-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK15-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK15-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK15-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK15-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK15-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK15-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK15-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK15-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK15-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK15-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK15-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK15-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK15-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK15-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP59]])
-// CHECK15-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK15-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK15:       omp_offload.failed15:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK15:       omp_offload.cont16:
-// CHECK15-NEXT:    [[TMP62:%.*]] = load i32, i32* [[M]], align 4
-// CHECK15-NEXT:    store i32 [[TMP62]], i32* [[M_CASTED]], align 4
-// CHECK15-NEXT:    [[TMP63:%.*]] = load i32, i32* [[M_CASTED]], align 4
-// CHECK15-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP64]], i32* [[N_CASTED17]], align 4
-// CHECK15-NEXT:    [[TMP65:%.*]] = load i32, i32* [[N_CASTED17]], align 4
-// CHECK15-NEXT:    [[TMP66:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK15-NEXT:    [[TMP67:%.*]] = sext i32 [[TMP66]] to i64
-// CHECK15-NEXT:    [[TMP68:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES21]] to i8*
-// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP68]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
-// CHECK15-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP63]], i32* [[TMP70]], align 4
-// CHECK15-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP63]], i32* [[TMP72]], align 4
-// CHECK15-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP73]], align 4
-// CHECK15-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP65]], i32* [[TMP75]], align 4
-// CHECK15-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP65]], i32* [[TMP77]], align 4
-// CHECK15-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP78]], align 4
-// CHECK15-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP80]], align 4
-// CHECK15-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP82]], align 4
-// CHECK15-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 2
-// CHECK15-NEXT:    store i8* null, i8** [[TMP83]], align 4
-// CHECK15-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 3
-// CHECK15-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP85]], align 4
-// CHECK15-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 3
-// CHECK15-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP87]], align 4
-// CHECK15-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 3
-// CHECK15-NEXT:    store i64 [[TMP67]], i64* [[TMP88]], align 4
-// CHECK15-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK15-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK15-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP67]])
+// CHECK15-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK15-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK15-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK15-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK15-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK15-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK15-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK15-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK15-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK15-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK15-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK15:       omp_offload.failed16:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK15:       omp_offload.cont17:
+// CHECK15-NEXT:    [[TMP78:%.*]] = load i32, i32* [[M]], align 4
+// CHECK15-NEXT:    store i32 [[TMP78]], i32* [[M_CASTED]], align 4
+// CHECK15-NEXT:    [[TMP79:%.*]] = load i32, i32* [[M_CASTED]], align 4
+// CHECK15-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP80]], i32* [[N_CASTED18]], align 4
+// CHECK15-NEXT:    [[TMP81:%.*]] = load i32, i32* [[N_CASTED18]], align 4
+// CHECK15-NEXT:    [[TMP82:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK15-NEXT:    [[TMP83:%.*]] = sext i32 [[TMP82]] to i64
+// CHECK15-NEXT:    [[TMP84:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
+// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP84]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
+// CHECK15-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP79]], i32* [[TMP86]], align 4
+// CHECK15-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP88:%.*]] = bitcast i8** [[TMP87]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP79]], i32* [[TMP88]], align 4
+// CHECK15-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
 // CHECK15-NEXT:    store i8* null, i8** [[TMP89]], align 4
-// CHECK15-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP93:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP93]], i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK15-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK15-NEXT:    [[SUB25:%.*]] = sub nsw i32 [[TMP94]], 0
-// CHECK15-NEXT:    [[DIV26:%.*]] = sdiv i32 [[SUB25]], 1
-// CHECK15-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[DIV26]], 1
-// CHECK15-NEXT:    store i32 [[SUB27]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK15-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK15-NEXT:    [[ADD28:%.*]] = add nsw i32 [[TMP95]], 1
-// CHECK15-NEXT:    [[TMP96:%.*]] = zext i32 [[ADD28]] to i64
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP96]])
-// CHECK15-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158.region_id, i32 4, i8** [[TMP90]], i8** [[TMP91]], i64* [[TMP92]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
-// CHECK15-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
-// CHECK15:       omp_offload.failed29:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158(i32 [[TMP63]], i32 [[TMP65]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
-// CHECK15:       omp_offload.cont30:
-// CHECK15-NEXT:    [[TMP99:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP99]], i32* [[N_CASTED31]], align 4
-// CHECK15-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N_CASTED31]], align 4
-// CHECK15-NEXT:    [[TMP101:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK15-NEXT:    [[TMP102:%.*]] = sext i32 [[TMP101]] to i64
-// CHECK15-NEXT:    [[TMP103:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES35]] to i8*
-// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP103]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
-// CHECK15-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP100]], i32* [[TMP105]], align 4
-// CHECK15-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP100]], i32* [[TMP107]], align 4
-// CHECK15-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP108]], align 4
-// CHECK15-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP110]], align 4
-// CHECK15-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP112]], align 4
-// CHECK15-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP113]], align 4
-// CHECK15-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP115]], align 4
-// CHECK15-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP117]], align 4
-// CHECK15-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES35]], i32 0, i32 2
-// CHECK15-NEXT:    store i64 [[TMP102]], i64* [[TMP118]], align 4
-// CHECK15-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 2
-// CHECK15-NEXT:    store i8* null, i8** [[TMP119]], align 4
-// CHECK15-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES35]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP81]], i32* [[TMP91]], align 4
+// CHECK15-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP81]], i32* [[TMP93]], align 4
+// CHECK15-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP94]], align 4
+// CHECK15-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP96]], align 4
+// CHECK15-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP98]], align 4
+// CHECK15-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
+// CHECK15-NEXT:    store i8* null, i8** [[TMP99]], align 4
+// CHECK15-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP101]], align 4
+// CHECK15-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP103]], align 4
+// CHECK15-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 3
+// CHECK15-NEXT:    store i64 [[TMP83]], i64* [[TMP104]], align 4
+// CHECK15-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK15-NEXT:    store i8* null, i8** [[TMP105]], align 4
+// CHECK15-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP109:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP109]], i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK15-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK15-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP110]], 0
+// CHECK15-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
+// CHECK15-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
+// CHECK15-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK15-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK15-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP111]], 1
+// CHECK15-NEXT:    [[TMP112:%.*]] = zext i32 [[ADD29]] to i64
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP112]])
+// CHECK15-NEXT:    [[KERNEL_ARGS30:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP113]], align 4
+// CHECK15-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 4, i32* [[TMP114]], align 4
+// CHECK15-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 4
+// CHECK15-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK15-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP108]], i64** [[TMP117]], align 4
+// CHECK15-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP118]], align 4
+// CHECK15-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP119]], align 4
+// CHECK15-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK15-NEXT:    [[TMP121:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]])
+// CHECK15-NEXT:    [[TMP122:%.*]] = icmp ne i32 [[TMP121]], 0
+// CHECK15-NEXT:    br i1 [[TMP122]], label [[OMP_OFFLOAD_FAILED31:%.*]], label [[OMP_OFFLOAD_CONT32:%.*]]
+// CHECK15:       omp_offload.failed31:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158(i32 [[TMP79]], i32 [[TMP81]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT32]]
+// CHECK15:       omp_offload.cont32:
 // CHECK15-NEXT:    [[TMP123:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP123]], i32* [[DOTCAPTURE_EXPR_37]], align 4
-// CHECK15-NEXT:    [[TMP124:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_37]], align 4
-// CHECK15-NEXT:    [[SUB39:%.*]] = sub nsw i32 [[TMP124]], 0
-// CHECK15-NEXT:    [[DIV40:%.*]] = sdiv i32 [[SUB39]], 1
-// CHECK15-NEXT:    [[SUB41:%.*]] = sub nsw i32 [[DIV40]], 1
-// CHECK15-NEXT:    store i32 [[SUB41]], i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK15-NEXT:    [[TMP125:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK15-NEXT:    [[ADD42:%.*]] = add nsw i32 [[TMP125]], 1
-// CHECK15-NEXT:    [[TMP126:%.*]] = zext i32 [[ADD42]] to i64
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP126]])
-// CHECK15-NEXT:    [[TMP127:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163.region_id, i32 3, i8** [[TMP120]], i8** [[TMP121]], i64* [[TMP122]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP128:%.*]] = icmp ne i32 [[TMP127]], 0
-// CHECK15-NEXT:    br i1 [[TMP128]], label [[OMP_OFFLOAD_FAILED43:%.*]], label [[OMP_OFFLOAD_CONT44:%.*]]
-// CHECK15:       omp_offload.failed43:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163(i32 [[TMP100]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT44]]
-// CHECK15:       omp_offload.cont44:
-// CHECK15-NEXT:    [[TMP129:%.*]] = load i32, i32* [[M]], align 4
-// CHECK15-NEXT:    store i32 [[TMP129]], i32* [[M_CASTED45]], align 4
-// CHECK15-NEXT:    [[TMP130:%.*]] = load i32, i32* [[M_CASTED45]], align 4
-// CHECK15-NEXT:    [[TMP131:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP131]], i32* [[N_CASTED46]], align 4
-// CHECK15-NEXT:    [[TMP132:%.*]] = load i32, i32* [[N_CASTED46]], align 4
-// CHECK15-NEXT:    [[TMP133:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK15-NEXT:    [[TMP134:%.*]] = sext i32 [[TMP133]] to i64
-// CHECK15-NEXT:    [[TMP135:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES50]] to i8*
-// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP135]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
-// CHECK15-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP130]], i32* [[TMP137]], align 4
-// CHECK15-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP130]], i32* [[TMP139]], align 4
-// CHECK15-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP140]], align 4
-// CHECK15-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP132]], i32* [[TMP142]], align 4
-// CHECK15-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP132]], i32* [[TMP144]], align 4
-// CHECK15-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP145]], align 4
-// CHECK15-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP147]], align 4
-// CHECK15-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 2
-// CHECK15-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP149]], align 4
-// CHECK15-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 2
-// CHECK15-NEXT:    store i8* null, i8** [[TMP150]], align 4
-// CHECK15-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 3
-// CHECK15-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP152]], align 4
-// CHECK15-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 3
-// CHECK15-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i32**
-// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP154]], align 4
-// CHECK15-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES50]], i32 0, i32 3
-// CHECK15-NEXT:    store i64 [[TMP134]], i64* [[TMP155]], align 4
-// CHECK15-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 3
-// CHECK15-NEXT:    store i8* null, i8** [[TMP156]], align 4
-// CHECK15-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES50]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP160:%.*]] = load i32, i32* [[N]], align 4
-// CHECK15-NEXT:    store i32 [[TMP160]], i32* [[DOTCAPTURE_EXPR_52]], align 4
-// CHECK15-NEXT:    [[TMP161:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_52]], align 4
-// CHECK15-NEXT:    [[SUB54:%.*]] = sub nsw i32 [[TMP161]], 0
-// CHECK15-NEXT:    [[DIV55:%.*]] = sdiv i32 [[SUB54]], 1
-// CHECK15-NEXT:    [[SUB56:%.*]] = sub nsw i32 [[DIV55]], 1
-// CHECK15-NEXT:    store i32 [[SUB56]], i32* [[DOTCAPTURE_EXPR_53]], align 4
-// CHECK15-NEXT:    [[TMP162:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_53]], align 4
-// CHECK15-NEXT:    [[ADD57:%.*]] = add nsw i32 [[TMP162]], 1
-// CHECK15-NEXT:    [[TMP163:%.*]] = zext i32 [[ADD57]] to i64
-// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP163]])
-// CHECK15-NEXT:    [[TMP164:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168.region_id, i32 4, i8** [[TMP157]], i8** [[TMP158]], i64* [[TMP159]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP165:%.*]] = icmp ne i32 [[TMP164]], 0
-// CHECK15-NEXT:    br i1 [[TMP165]], label [[OMP_OFFLOAD_FAILED58:%.*]], label [[OMP_OFFLOAD_CONT59:%.*]]
-// CHECK15:       omp_offload.failed58:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168(i32 [[TMP130]], i32 [[TMP132]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT59]]
-// CHECK15:       omp_offload.cont59:
-// CHECK15-NEXT:    [[TMP166:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK15-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP166]])
+// CHECK15-NEXT:    store i32 [[TMP123]], i32* [[N_CASTED33]], align 4
+// CHECK15-NEXT:    [[TMP124:%.*]] = load i32, i32* [[N_CASTED33]], align 4
+// CHECK15-NEXT:    [[TMP125:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK15-NEXT:    [[TMP126:%.*]] = sext i32 [[TMP125]] to i64
+// CHECK15-NEXT:    [[TMP127:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES37]] to i8*
+// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP127]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
+// CHECK15-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP124]], i32* [[TMP129]], align 4
+// CHECK15-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP124]], i32* [[TMP131]], align 4
+// CHECK15-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP132]], align 4
+// CHECK15-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP134]], align 4
+// CHECK15-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP136]], align 4
+// CHECK15-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP137]], align 4
+// CHECK15-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP139]], align 4
+// CHECK15-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP141]], align 4
+// CHECK15-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES37]], i32 0, i32 2
+// CHECK15-NEXT:    store i64 [[TMP126]], i64* [[TMP142]], align 4
+// CHECK15-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i32 0, i32 2
+// CHECK15-NEXT:    store i8* null, i8** [[TMP143]], align 4
+// CHECK15-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES37]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP147:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP147]], i32* [[DOTCAPTURE_EXPR_39]], align 4
+// CHECK15-NEXT:    [[TMP148:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_39]], align 4
+// CHECK15-NEXT:    [[SUB41:%.*]] = sub nsw i32 [[TMP148]], 0
+// CHECK15-NEXT:    [[DIV42:%.*]] = sdiv i32 [[SUB41]], 1
+// CHECK15-NEXT:    [[SUB43:%.*]] = sub nsw i32 [[DIV42]], 1
+// CHECK15-NEXT:    store i32 [[SUB43]], i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK15-NEXT:    [[TMP149:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK15-NEXT:    [[ADD44:%.*]] = add nsw i32 [[TMP149]], 1
+// CHECK15-NEXT:    [[TMP150:%.*]] = zext i32 [[ADD44]] to i64
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP150]])
+// CHECK15-NEXT:    [[KERNEL_ARGS45:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP151]], align 4
+// CHECK15-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 3, i32* [[TMP152]], align 4
+// CHECK15-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP144]], i8*** [[TMP153]], align 4
+// CHECK15-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP145]], i8*** [[TMP154]], align 4
+// CHECK15-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP146]], i64** [[TMP155]], align 4
+// CHECK15-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP156]], align 4
+// CHECK15-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP157]], align 4
+// CHECK15-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP158]], align 4
+// CHECK15-NEXT:    [[TMP159:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]])
+// CHECK15-NEXT:    [[TMP160:%.*]] = icmp ne i32 [[TMP159]], 0
+// CHECK15-NEXT:    br i1 [[TMP160]], label [[OMP_OFFLOAD_FAILED46:%.*]], label [[OMP_OFFLOAD_CONT47:%.*]]
+// CHECK15:       omp_offload.failed46:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163(i32 [[TMP124]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT47]]
+// CHECK15:       omp_offload.cont47:
+// CHECK15-NEXT:    [[TMP161:%.*]] = load i32, i32* [[M]], align 4
+// CHECK15-NEXT:    store i32 [[TMP161]], i32* [[M_CASTED48]], align 4
+// CHECK15-NEXT:    [[TMP162:%.*]] = load i32, i32* [[M_CASTED48]], align 4
+// CHECK15-NEXT:    [[TMP163:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP163]], i32* [[N_CASTED49]], align 4
+// CHECK15-NEXT:    [[TMP164:%.*]] = load i32, i32* [[N_CASTED49]], align 4
+// CHECK15-NEXT:    [[TMP165:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK15-NEXT:    [[TMP166:%.*]] = sext i32 [[TMP165]] to i64
+// CHECK15-NEXT:    [[TMP167:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES53]] to i8*
+// CHECK15-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP167]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
+// CHECK15-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP162]], i32* [[TMP169]], align 4
+// CHECK15-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP171:%.*]] = bitcast i8** [[TMP170]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP162]], i32* [[TMP171]], align 4
+// CHECK15-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP172]], align 4
+// CHECK15-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP174:%.*]] = bitcast i8** [[TMP173]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP164]], i32* [[TMP174]], align 4
+// CHECK15-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP176:%.*]] = bitcast i8** [[TMP175]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP164]], i32* [[TMP176]], align 4
+// CHECK15-NEXT:    [[TMP177:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP177]], align 4
+// CHECK15-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP179]], align 4
+// CHECK15-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 2
+// CHECK15-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP0]], i32* [[TMP181]], align 4
+// CHECK15-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 2
+// CHECK15-NEXT:    store i8* null, i8** [[TMP182]], align 4
+// CHECK15-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP184:%.*]] = bitcast i8** [[TMP183]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP184]], align 4
+// CHECK15-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 3
+// CHECK15-NEXT:    [[TMP186:%.*]] = bitcast i8** [[TMP185]] to i32**
+// CHECK15-NEXT:    store i32* [[VLA]], i32** [[TMP186]], align 4
+// CHECK15-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES53]], i32 0, i32 3
+// CHECK15-NEXT:    store i64 [[TMP166]], i64* [[TMP187]], align 4
+// CHECK15-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 3
+// CHECK15-NEXT:    store i8* null, i8** [[TMP188]], align 4
+// CHECK15-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES53]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP192:%.*]] = load i32, i32* [[N]], align 4
+// CHECK15-NEXT:    store i32 [[TMP192]], i32* [[DOTCAPTURE_EXPR_55]], align 4
+// CHECK15-NEXT:    [[TMP193:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_55]], align 4
+// CHECK15-NEXT:    [[SUB57:%.*]] = sub nsw i32 [[TMP193]], 0
+// CHECK15-NEXT:    [[DIV58:%.*]] = sdiv i32 [[SUB57]], 1
+// CHECK15-NEXT:    [[SUB59:%.*]] = sub nsw i32 [[DIV58]], 1
+// CHECK15-NEXT:    store i32 [[SUB59]], i32* [[DOTCAPTURE_EXPR_56]], align 4
+// CHECK15-NEXT:    [[TMP194:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_56]], align 4
+// CHECK15-NEXT:    [[ADD60:%.*]] = add nsw i32 [[TMP194]], 1
+// CHECK15-NEXT:    [[TMP195:%.*]] = zext i32 [[ADD60]] to i64
+// CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP195]])
+// CHECK15-NEXT:    [[KERNEL_ARGS61:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP196]], align 4
+// CHECK15-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 4, i32* [[TMP197]], align 4
+// CHECK15-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP189]], i8*** [[TMP198]], align 4
+// CHECK15-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP190]], i8*** [[TMP199]], align 4
+// CHECK15-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* [[TMP191]], i64** [[TMP200]], align 4
+// CHECK15-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP201]], align 4
+// CHECK15-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP202]], align 4
+// CHECK15-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP203]], align 4
+// CHECK15-NEXT:    [[TMP204:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]])
+// CHECK15-NEXT:    [[TMP205:%.*]] = icmp ne i32 [[TMP204]], 0
+// CHECK15-NEXT:    br i1 [[TMP205]], label [[OMP_OFFLOAD_FAILED62:%.*]], label [[OMP_OFFLOAD_CONT63:%.*]]
+// CHECK15:       omp_offload.failed62:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168(i32 [[TMP162]], i32 [[TMP164]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT63]]
+// CHECK15:       omp_offload.cont63:
+// CHECK15-NEXT:    [[TMP206:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK15-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP206]])
 // CHECK15-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK15-NEXT:    [[TMP167:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK15-NEXT:    call void @llvm.stackrestore(i8* [[TMP167]])
-// CHECK15-NEXT:    [[TMP168:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK15-NEXT:    ret i32 [[TMP168]]
+// CHECK15-NEXT:    [[TMP207:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK15-NEXT:    call void @llvm.stackrestore(i8* [[TMP207]])
+// CHECK15-NEXT:    [[TMP208:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK15-NEXT:    ret i32 [[TMP208]]
 //
 //
 // CHECK15-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148
@@ -7725,19 +8320,19 @@ int main (int argc, char **argv) {
 // CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 4
 // CHECK15-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    [[M_CASTED:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 4
-// CHECK15-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[M_CASTED19:%.*]] = alloca i32, align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [2 x i8*], align 4
-// CHECK15-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 4
+// CHECK15-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[M_CASTED22:%.*]] = alloca i32, align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [2 x i8*], align 4
+// CHECK15-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
 // CHECK15-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK15-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK15-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -7751,107 +8346,192 @@ int main (int argc, char **argv) {
 // CHECK15-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK15-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK15-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK15-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK15-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK15-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK15-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK15-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK15-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK15-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK15-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK15-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK15-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK15-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK15-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK15:       omp_offload.failed:
 // CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK15:       omp_offload.cont:
-// CHECK15-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK15-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK15-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK15-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK15-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK15-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK15-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l121.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK15-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK15:       omp_offload.failed5:
+// CHECK15-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK15-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK15-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK15-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK15-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK15-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK15-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK15-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK15-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l121.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK15-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK15-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK15:       omp_offload.failed6:
 // CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l121([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK15:       omp_offload.cont6:
-// CHECK15-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK15-NEXT:    store i32 [[TMP18]], i32* [[M_CASTED]], align 4
-// CHECK15-NEXT:    [[TMP19:%.*]] = load i32, i32* [[M_CASTED]], align 4
-// CHECK15-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP19]], i32* [[TMP21]], align 4
-// CHECK15-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP19]], i32* [[TMP23]], align 4
-// CHECK15-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP24]], align 4
-// CHECK15-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP26]], align 4
-// CHECK15-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP28]], align 4
-// CHECK15-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP29]], align 4
-// CHECK15-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK15:       omp_offload.cont7:
+// CHECK15-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK15-NEXT:    store i32 [[TMP34]], i32* [[M_CASTED]], align 4
+// CHECK15-NEXT:    [[TMP35:%.*]] = load i32, i32* [[M_CASTED]], align 4
+// CHECK15-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP35]], i32* [[TMP37]], align 4
+// CHECK15-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP35]], i32* [[TMP39]], align 4
+// CHECK15-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP40]], align 4
+// CHECK15-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP42]], align 4
+// CHECK15-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP44]], align 4
+// CHECK15-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP45]], align 4
+// CHECK15-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126.region_id, i32 2, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK15-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK15:       omp_offload.failed11:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126(i32 [[TMP19]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK15:       omp_offload.cont12:
-// CHECK15-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 4
-// CHECK15-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 4
-// CHECK15-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP38]], align 4
-// CHECK15-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK15-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK15-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 2, i32* [[TMP49]], align 4
+// CHECK15-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP46]], i8*** [[TMP50]], align 4
+// CHECK15-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 4
+// CHECK15-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP52]], align 4
+// CHECK15-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP53]], align 4
+// CHECK15-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP54]], align 4
+// CHECK15-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP55]], align 4
+// CHECK15-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK15-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK15-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK15:       omp_offload.failed13:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126(i32 [[TMP35]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK15:       omp_offload.cont14:
+// CHECK15-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP59]], align 4
+// CHECK15-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP61]], align 4
+// CHECK15-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP62]], align 4
+// CHECK15-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP41:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l131.region_id, i32 1, i8** [[TMP39]], i8** [[TMP40]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP42:%.*]] = icmp ne i32 [[TMP41]], 0
-// CHECK15-NEXT:    br i1 [[TMP42]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK15:       omp_offload.failed17:
+// CHECK15-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP65]], align 4
+// CHECK15-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK15-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP63]], i8*** [[TMP67]], align 4
+// CHECK15-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK15-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP69]], align 4
+// CHECK15-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK15-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP71]], align 4
+// CHECK15-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK15-NEXT:    [[TMP73:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l131.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK15-NEXT:    [[TMP74:%.*]] = icmp ne i32 [[TMP73]], 0
+// CHECK15-NEXT:    br i1 [[TMP74]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK15:       omp_offload.failed20:
 // CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l131([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK15:       omp_offload.cont18:
-// CHECK15-NEXT:    [[TMP43:%.*]] = load i32, i32* [[M]], align 4
-// CHECK15-NEXT:    store i32 [[TMP43]], i32* [[M_CASTED19]], align 4
-// CHECK15-NEXT:    [[TMP44:%.*]] = load i32, i32* [[M_CASTED19]], align 4
-// CHECK15-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP44]], i32* [[TMP46]], align 4
-// CHECK15-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
-// CHECK15-NEXT:    store i32 [[TMP44]], i32* [[TMP48]], align 4
-// CHECK15-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
-// CHECK15-NEXT:    store i8* null, i8** [[TMP49]], align 4
-// CHECK15-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP51]], align 4
-// CHECK15-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK15-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to [10 x i32]**
-// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP53]], align 4
-// CHECK15-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
-// CHECK15-NEXT:    store i8* null, i8** [[TMP54]], align 4
-// CHECK15-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK15-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK15:       omp_offload.cont21:
+// CHECK15-NEXT:    [[TMP75:%.*]] = load i32, i32* [[M]], align 4
+// CHECK15-NEXT:    store i32 [[TMP75]], i32* [[M_CASTED22]], align 4
+// CHECK15-NEXT:    [[TMP76:%.*]] = load i32, i32* [[M_CASTED22]], align 4
+// CHECK15-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP76]], i32* [[TMP78]], align 4
+// CHECK15-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32*
+// CHECK15-NEXT:    store i32 [[TMP76]], i32* [[TMP80]], align 4
+// CHECK15-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 0
+// CHECK15-NEXT:    store i8* null, i8** [[TMP81]], align 4
+// CHECK15-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP83]], align 4
+// CHECK15-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
+// CHECK15-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to [10 x i32]**
+// CHECK15-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP85]], align 4
+// CHECK15-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 1
+// CHECK15-NEXT:    store i8* null, i8** [[TMP86]], align 4
+// CHECK15-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK15-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
 // CHECK15-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK15-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK15-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
-// CHECK15-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
-// CHECK15:       omp_offload.failed24:
-// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136(i32 [[TMP44]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
-// CHECK15:       omp_offload.cont25:
+// CHECK15-NEXT:    [[KERNEL_ARGS27:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK15-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 0
+// CHECK15-NEXT:    store i32 1, i32* [[TMP89]], align 4
+// CHECK15-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 1
+// CHECK15-NEXT:    store i32 2, i32* [[TMP90]], align 4
+// CHECK15-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 2
+// CHECK15-NEXT:    store i8** [[TMP87]], i8*** [[TMP91]], align 4
+// CHECK15-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 3
+// CHECK15-NEXT:    store i8** [[TMP88]], i8*** [[TMP92]], align 4
+// CHECK15-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 4
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP93]], align 4
+// CHECK15-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 5
+// CHECK15-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP94]], align 4
+// CHECK15-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 6
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP95]], align 4
+// CHECK15-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 7
+// CHECK15-NEXT:    store i8** null, i8*** [[TMP96]], align 4
+// CHECK15-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]])
+// CHECK15-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
+// CHECK15-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
+// CHECK15:       omp_offload.failed28:
+// CHECK15-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136(i32 [[TMP76]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK15-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
+// CHECK15:       omp_offload.cont29:
 // CHECK15-NEXT:    ret i32 0
 //
 //
@@ -8665,31 +9345,31 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[M_CASTED:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES24:%.*]] = alloca [4 x i64], align 8
-// CHECK17-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
+// CHECK17-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[N_CASTED34:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS36:%.*]] = alloca [3 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS37:%.*]] = alloca [3 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS38:%.*]] = alloca [3 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES39:%.*]] = alloca [3 x i64], align 8
-// CHECK17-NEXT:    [[_TMP40:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_41:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_42:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[M_CASTED49:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[N_CASTED51:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS53:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS54:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS55:%.*]] = alloca [4 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES56:%.*]] = alloca [4 x i64], align 8
-// CHECK17-NEXT:    [[_TMP57:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_58:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_59:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[N_CASTED36:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS38:%.*]] = alloca [3 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS39:%.*]] = alloca [3 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS40:%.*]] = alloca [3 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES41:%.*]] = alloca [3 x i64], align 8
+// CHECK17-NEXT:    [[_TMP42:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_43:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_44:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[M_CASTED52:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[N_CASTED54:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS56:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS57:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS58:%.*]] = alloca [4 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES59:%.*]] = alloca [4 x i64], align 8
+// CHECK17-NEXT:    [[_TMP60:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_61:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_62:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK17-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK17-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -8748,260 +9428,345 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK17-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK17-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK17-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK17-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK17-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK17-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK17-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK17-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK17-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK17-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
-// CHECK17-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK17-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK17-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK17-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK17-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
-// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK17-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK17-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK17-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK17-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK17-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK17-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK17-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK17-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK17-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK17-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK17-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK17-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK17-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
+// CHECK17-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK17-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK17-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK17-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK17-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK17-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK17-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK17-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK17-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK17-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK17-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK17-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK17-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK17-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK17-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK17-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK17-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP58]])
-// CHECK17-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK17-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK17:       omp_offload.failed16:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK17:       omp_offload.cont17:
-// CHECK17-NEXT:    [[TMP61:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    [[CONV18:%.*]] = bitcast i64* [[M_CASTED]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP61]], i32* [[CONV18]], align 4
-// CHECK17-NEXT:    [[TMP62:%.*]] = load i64, i64* [[M_CASTED]], align 8
-// CHECK17-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP63]], i32* [[CONV20]], align 4
-// CHECK17-NEXT:    [[TMP64:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK17-NEXT:    [[TMP65:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK17-NEXT:    [[TMP66:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES24]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP66]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
-// CHECK17-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP62]], i64* [[TMP68]], align 8
-// CHECK17-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP62]], i64* [[TMP70]], align 8
-// CHECK17-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP71]], align 8
-// CHECK17-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP64]], i64* [[TMP73]], align 8
-// CHECK17-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP64]], i64* [[TMP75]], align 8
-// CHECK17-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP76]], align 8
-// CHECK17-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP78]], align 8
-// CHECK17-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP80]], align 8
-// CHECK17-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP81]], align 8
-// CHECK17-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 8
-// CHECK17-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP85]], align 8
-// CHECK17-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 3
-// CHECK17-NEXT:    store i64 [[TMP65]], i64* [[TMP86]], align 8
-// CHECK17-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK17-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK17-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK17-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP66]])
+// CHECK17-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK17-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK17-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK17-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK17-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK17-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK17-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK17-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK17-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK17-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK17-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK17:       omp_offload.failed17:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK17:       omp_offload.cont18:
+// CHECK17-NEXT:    [[TMP77:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    [[CONV19:%.*]] = bitcast i64* [[M_CASTED]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP77]], i32* [[CONV19]], align 4
+// CHECK17-NEXT:    [[TMP78:%.*]] = load i64, i64* [[M_CASTED]], align 8
+// CHECK17-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP79]], i32* [[CONV21]], align 4
+// CHECK17-NEXT:    [[TMP80:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK17-NEXT:    [[TMP81:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK17-NEXT:    [[TMP82:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP82]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
+// CHECK17-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP78]], i64* [[TMP84]], align 8
+// CHECK17-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP78]], i64* [[TMP86]], align 8
+// CHECK17-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
 // CHECK17-NEXT:    store i8* null, i8** [[TMP87]], align 8
-// CHECK17-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP91:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP91]], i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK17-NEXT:    [[TMP92:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK17-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[TMP92]], 0
-// CHECK17-NEXT:    [[DIV29:%.*]] = sdiv i32 [[SUB28]], 1
-// CHECK17-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[DIV29]], 1
-// CHECK17-NEXT:    store i32 [[SUB30]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK17-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK17-NEXT:    [[ADD31:%.*]] = add nsw i32 [[TMP93]], 1
-// CHECK17-NEXT:    [[TMP94:%.*]] = zext i32 [[ADD31]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP94]])
-// CHECK17-NEXT:    [[TMP95:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158.region_id, i32 4, i8** [[TMP88]], i8** [[TMP89]], i64* [[TMP90]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP96:%.*]] = icmp ne i32 [[TMP95]], 0
-// CHECK17-NEXT:    br i1 [[TMP96]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
-// CHECK17:       omp_offload.failed32:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158(i64 [[TMP62]], i64 [[TMP64]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
-// CHECK17:       omp_offload.cont33:
-// CHECK17-NEXT:    [[TMP97:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    [[CONV35:%.*]] = bitcast i64* [[N_CASTED34]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP97]], i32* [[CONV35]], align 4
-// CHECK17-NEXT:    [[TMP98:%.*]] = load i64, i64* [[N_CASTED34]], align 8
-// CHECK17-NEXT:    [[TMP99:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK17-NEXT:    [[TMP100:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES39]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP100]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
-// CHECK17-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP98]], i64* [[TMP102]], align 8
-// CHECK17-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP98]], i64* [[TMP104]], align 8
-// CHECK17-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP105]], align 8
-// CHECK17-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP107]], align 8
-// CHECK17-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP109]], align 8
-// CHECK17-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP110]], align 8
-// CHECK17-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP112]], align 8
-// CHECK17-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP114]], align 8
-// CHECK17-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES39]], i32 0, i32 2
-// CHECK17-NEXT:    store i64 [[TMP99]], i64* [[TMP115]], align 8
-// CHECK17-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP116]], align 8
-// CHECK17-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES39]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP120:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP120]], i32* [[DOTCAPTURE_EXPR_41]], align 4
-// CHECK17-NEXT:    [[TMP121:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_41]], align 4
-// CHECK17-NEXT:    [[SUB43:%.*]] = sub nsw i32 [[TMP121]], 0
-// CHECK17-NEXT:    [[DIV44:%.*]] = sdiv i32 [[SUB43]], 1
-// CHECK17-NEXT:    [[SUB45:%.*]] = sub nsw i32 [[DIV44]], 1
-// CHECK17-NEXT:    store i32 [[SUB45]], i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK17-NEXT:    [[TMP122:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK17-NEXT:    [[ADD46:%.*]] = add nsw i32 [[TMP122]], 1
-// CHECK17-NEXT:    [[TMP123:%.*]] = zext i32 [[ADD46]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP123]])
-// CHECK17-NEXT:    [[TMP124:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163.region_id, i32 3, i8** [[TMP117]], i8** [[TMP118]], i64* [[TMP119]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP125:%.*]] = icmp ne i32 [[TMP124]], 0
-// CHECK17-NEXT:    br i1 [[TMP125]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
-// CHECK17:       omp_offload.failed47:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163(i64 [[TMP98]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
-// CHECK17:       omp_offload.cont48:
-// CHECK17-NEXT:    [[TMP126:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    [[CONV50:%.*]] = bitcast i64* [[M_CASTED49]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP126]], i32* [[CONV50]], align 4
-// CHECK17-NEXT:    [[TMP127:%.*]] = load i64, i64* [[M_CASTED49]], align 8
-// CHECK17-NEXT:    [[TMP128:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    [[CONV52:%.*]] = bitcast i64* [[N_CASTED51]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP128]], i32* [[CONV52]], align 4
-// CHECK17-NEXT:    [[TMP129:%.*]] = load i64, i64* [[N_CASTED51]], align 8
-// CHECK17-NEXT:    [[TMP130:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK17-NEXT:    [[TMP131:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES56]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP131]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
-// CHECK17-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP80]], i64* [[TMP89]], align 8
+// CHECK17-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP80]], i64* [[TMP91]], align 8
+// CHECK17-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP92]], align 8
+// CHECK17-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP94]], align 8
+// CHECK17-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP96]], align 8
+// CHECK17-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP97]], align 8
+// CHECK17-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 8
+// CHECK17-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP101]], align 8
+// CHECK17-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 3
+// CHECK17-NEXT:    store i64 [[TMP81]], i64* [[TMP102]], align 8
+// CHECK17-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK17-NEXT:    store i8* null, i8** [[TMP103]], align 8
+// CHECK17-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP107:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP107]], i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK17-NEXT:    [[TMP108:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK17-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP108]], 0
+// CHECK17-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
+// CHECK17-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
+// CHECK17-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK17-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK17-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP109]], 1
+// CHECK17-NEXT:    [[TMP110:%.*]] = zext i32 [[ADD32]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP110]])
+// CHECK17-NEXT:    [[KERNEL_ARGS33:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP111]], align 4
+// CHECK17-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 4, i32* [[TMP112]], align 4
+// CHECK17-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP104]], i8*** [[TMP113]], align 8
+// CHECK17-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK17-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP106]], i64** [[TMP115]], align 8
+// CHECK17-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP116]], align 8
+// CHECK17-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP117]], align 8
+// CHECK17-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK17-NEXT:    [[TMP119:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]])
+// CHECK17-NEXT:    [[TMP120:%.*]] = icmp ne i32 [[TMP119]], 0
+// CHECK17-NEXT:    br i1 [[TMP120]], label [[OMP_OFFLOAD_FAILED34:%.*]], label [[OMP_OFFLOAD_CONT35:%.*]]
+// CHECK17:       omp_offload.failed34:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158(i64 [[TMP78]], i64 [[TMP80]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT35]]
+// CHECK17:       omp_offload.cont35:
+// CHECK17-NEXT:    [[TMP121:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    [[CONV37:%.*]] = bitcast i64* [[N_CASTED36]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP121]], i32* [[CONV37]], align 4
+// CHECK17-NEXT:    [[TMP122:%.*]] = load i64, i64* [[N_CASTED36]], align 8
+// CHECK17-NEXT:    [[TMP123:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK17-NEXT:    [[TMP124:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES41]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP124]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
+// CHECK17-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP122]], i64* [[TMP126]], align 8
+// CHECK17-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP122]], i64* [[TMP128]], align 8
+// CHECK17-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS40]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP129]], align 8
+// CHECK17-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP131]], align 8
+// CHECK17-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 1
 // CHECK17-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP127]], i64* [[TMP133]], align 8
-// CHECK17-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP127]], i64* [[TMP135]], align 8
-// CHECK17-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP136]], align 8
-// CHECK17-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP129]], i64* [[TMP138]], align 8
-// CHECK17-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP129]], i64* [[TMP140]], align 8
-// CHECK17-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP141]], align 8
-// CHECK17-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP143]], align 8
-// CHECK17-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP145:%.*]] = bitcast i8** [[TMP144]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP145]], align 8
-// CHECK17-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP146]], align 8
-// CHECK17-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP148]], align 8
-// CHECK17-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP150:%.*]] = bitcast i8** [[TMP149]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP150]], align 8
-// CHECK17-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES56]], i32 0, i32 3
-// CHECK17-NEXT:    store i64 [[TMP130]], i64* [[TMP151]], align 8
-// CHECK17-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 3
-// CHECK17-NEXT:    store i8* null, i8** [[TMP152]], align 8
-// CHECK17-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES56]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP156:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP156]], i32* [[DOTCAPTURE_EXPR_58]], align 4
-// CHECK17-NEXT:    [[TMP157:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_58]], align 4
-// CHECK17-NEXT:    [[SUB60:%.*]] = sub nsw i32 [[TMP157]], 0
-// CHECK17-NEXT:    [[DIV61:%.*]] = sdiv i32 [[SUB60]], 1
-// CHECK17-NEXT:    [[SUB62:%.*]] = sub nsw i32 [[DIV61]], 1
-// CHECK17-NEXT:    store i32 [[SUB62]], i32* [[DOTCAPTURE_EXPR_59]], align 4
-// CHECK17-NEXT:    [[TMP158:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_59]], align 4
-// CHECK17-NEXT:    [[ADD63:%.*]] = add nsw i32 [[TMP158]], 1
-// CHECK17-NEXT:    [[TMP159:%.*]] = zext i32 [[ADD63]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP159]])
-// CHECK17-NEXT:    [[TMP160:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168.region_id, i32 4, i8** [[TMP153]], i8** [[TMP154]], i64* [[TMP155]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP161:%.*]] = icmp ne i32 [[TMP160]], 0
-// CHECK17-NEXT:    br i1 [[TMP161]], label [[OMP_OFFLOAD_FAILED64:%.*]], label [[OMP_OFFLOAD_CONT65:%.*]]
-// CHECK17:       omp_offload.failed64:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168(i64 [[TMP127]], i64 [[TMP129]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT65]]
-// CHECK17:       omp_offload.cont65:
-// CHECK17-NEXT:    [[TMP162:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK17-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP162]])
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP133]], align 8
+// CHECK17-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS40]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP134]], align 8
+// CHECK17-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP136]], align 8
+// CHECK17-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP138]], align 8
+// CHECK17-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES41]], i32 0, i32 2
+// CHECK17-NEXT:    store i64 [[TMP123]], i64* [[TMP139]], align 8
+// CHECK17-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS40]], i64 0, i64 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP140]], align 8
+// CHECK17-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES41]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP144:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP144]], i32* [[DOTCAPTURE_EXPR_43]], align 4
+// CHECK17-NEXT:    [[TMP145:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_43]], align 4
+// CHECK17-NEXT:    [[SUB45:%.*]] = sub nsw i32 [[TMP145]], 0
+// CHECK17-NEXT:    [[DIV46:%.*]] = sdiv i32 [[SUB45]], 1
+// CHECK17-NEXT:    [[SUB47:%.*]] = sub nsw i32 [[DIV46]], 1
+// CHECK17-NEXT:    store i32 [[SUB47]], i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK17-NEXT:    [[TMP146:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK17-NEXT:    [[ADD48:%.*]] = add nsw i32 [[TMP146]], 1
+// CHECK17-NEXT:    [[TMP147:%.*]] = zext i32 [[ADD48]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP147]])
+// CHECK17-NEXT:    [[KERNEL_ARGS49:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP148]], align 4
+// CHECK17-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP149]], align 4
+// CHECK17-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP141]], i8*** [[TMP150]], align 8
+// CHECK17-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP142]], i8*** [[TMP151]], align 8
+// CHECK17-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP143]], i64** [[TMP152]], align 8
+// CHECK17-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP153]], align 8
+// CHECK17-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP154]], align 8
+// CHECK17-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP155]], align 8
+// CHECK17-NEXT:    [[TMP156:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]])
+// CHECK17-NEXT:    [[TMP157:%.*]] = icmp ne i32 [[TMP156]], 0
+// CHECK17-NEXT:    br i1 [[TMP157]], label [[OMP_OFFLOAD_FAILED50:%.*]], label [[OMP_OFFLOAD_CONT51:%.*]]
+// CHECK17:       omp_offload.failed50:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163(i64 [[TMP122]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT51]]
+// CHECK17:       omp_offload.cont51:
+// CHECK17-NEXT:    [[TMP158:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    [[CONV53:%.*]] = bitcast i64* [[M_CASTED52]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP158]], i32* [[CONV53]], align 4
+// CHECK17-NEXT:    [[TMP159:%.*]] = load i64, i64* [[M_CASTED52]], align 8
+// CHECK17-NEXT:    [[TMP160:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    [[CONV55:%.*]] = bitcast i64* [[N_CASTED54]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP160]], i32* [[CONV55]], align 4
+// CHECK17-NEXT:    [[TMP161:%.*]] = load i64, i64* [[N_CASTED54]], align 8
+// CHECK17-NEXT:    [[TMP162:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK17-NEXT:    [[TMP163:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES59]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP163]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
+// CHECK17-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP165:%.*]] = bitcast i8** [[TMP164]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP159]], i64* [[TMP165]], align 8
+// CHECK17-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP159]], i64* [[TMP167]], align 8
+// CHECK17-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP168]], align 8
+// CHECK17-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP170:%.*]] = bitcast i8** [[TMP169]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP161]], i64* [[TMP170]], align 8
+// CHECK17-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP172:%.*]] = bitcast i8** [[TMP171]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP161]], i64* [[TMP172]], align 8
+// CHECK17-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP173]], align 8
+// CHECK17-NEXT:    [[TMP174:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP175:%.*]] = bitcast i8** [[TMP174]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP175]], align 8
+// CHECK17-NEXT:    [[TMP176:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP177:%.*]] = bitcast i8** [[TMP176]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP1]], i64* [[TMP177]], align 8
+// CHECK17-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP178]], align 8
+// CHECK17-NEXT:    [[TMP179:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP180:%.*]] = bitcast i8** [[TMP179]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP180]], align 8
+// CHECK17-NEXT:    [[TMP181:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP182:%.*]] = bitcast i8** [[TMP181]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP182]], align 8
+// CHECK17-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES59]], i32 0, i32 3
+// CHECK17-NEXT:    store i64 [[TMP162]], i64* [[TMP183]], align 8
+// CHECK17-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 3
+// CHECK17-NEXT:    store i8* null, i8** [[TMP184]], align 8
+// CHECK17-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES59]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP188:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP188]], i32* [[DOTCAPTURE_EXPR_61]], align 4
+// CHECK17-NEXT:    [[TMP189:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_61]], align 4
+// CHECK17-NEXT:    [[SUB63:%.*]] = sub nsw i32 [[TMP189]], 0
+// CHECK17-NEXT:    [[DIV64:%.*]] = sdiv i32 [[SUB63]], 1
+// CHECK17-NEXT:    [[SUB65:%.*]] = sub nsw i32 [[DIV64]], 1
+// CHECK17-NEXT:    store i32 [[SUB65]], i32* [[DOTCAPTURE_EXPR_62]], align 4
+// CHECK17-NEXT:    [[TMP190:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_62]], align 4
+// CHECK17-NEXT:    [[ADD66:%.*]] = add nsw i32 [[TMP190]], 1
+// CHECK17-NEXT:    [[TMP191:%.*]] = zext i32 [[ADD66]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP191]])
+// CHECK17-NEXT:    [[KERNEL_ARGS67:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP192]], align 4
+// CHECK17-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 4, i32* [[TMP193]], align 4
+// CHECK17-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP185]], i8*** [[TMP194]], align 8
+// CHECK17-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP186]], i8*** [[TMP195]], align 8
+// CHECK17-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP187]], i64** [[TMP196]], align 8
+// CHECK17-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP197]], align 8
+// CHECK17-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP198]], align 8
+// CHECK17-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP199]], align 8
+// CHECK17-NEXT:    [[TMP200:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]])
+// CHECK17-NEXT:    [[TMP201:%.*]] = icmp ne i32 [[TMP200]], 0
+// CHECK17-NEXT:    br i1 [[TMP201]], label [[OMP_OFFLOAD_FAILED68:%.*]], label [[OMP_OFFLOAD_CONT69:%.*]]
+// CHECK17:       omp_offload.failed68:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168(i64 [[TMP159]], i64 [[TMP161]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT69]]
+// CHECK17:       omp_offload.cont69:
+// CHECK17-NEXT:    [[TMP202:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK17-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP202]])
 // CHECK17-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK17-NEXT:    [[TMP163:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP163]])
-// CHECK17-NEXT:    [[TMP164:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK17-NEXT:    ret i32 [[TMP164]]
+// CHECK17-NEXT:    [[TMP203:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP203]])
+// CHECK17-NEXT:    [[TMP204:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK17-NEXT:    ret i32 [[TMP204]]
 //
 //
 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148
@@ -10141,19 +10906,19 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 8
 // CHECK17-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[M_CASTED:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 8
-// CHECK17-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[M_CASTED19:%.*]] = alloca i64, align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [2 x i8*], align 8
-// CHECK17-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 8
+// CHECK17-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[M_CASTED22:%.*]] = alloca i64, align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [2 x i8*], align 8
+// CHECK17-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK17-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK17-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -10167,109 +10932,194 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK17-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK17-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK17-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK17-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK17-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK17-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK17-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK17-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
-// CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK17-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK17-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK17-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK17-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l121.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK17-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK17:       omp_offload.failed5:
+// CHECK17-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK17-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK17-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK17-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l121.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK17-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK17-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK17:       omp_offload.failed6:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l121([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK17:       omp_offload.cont6:
-// CHECK17-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK17:       omp_offload.cont7:
+// CHECK17-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
 // CHECK17-NEXT:    [[CONV:%.*]] = bitcast i64* [[M_CASTED]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP18]], i32* [[CONV]], align 4
-// CHECK17-NEXT:    [[TMP19:%.*]] = load i64, i64* [[M_CASTED]], align 8
-// CHECK17-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP19]], i64* [[TMP21]], align 8
-// CHECK17-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP19]], i64* [[TMP23]], align 8
-// CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP24]], align 8
-// CHECK17-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP26]], align 8
-// CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP28]], align 8
-// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP29]], align 8
-// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 [[TMP34]], i32* [[CONV]], align 4
+// CHECK17-NEXT:    [[TMP35:%.*]] = load i64, i64* [[M_CASTED]], align 8
+// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP35]], i64* [[TMP37]], align 8
+// CHECK17-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP35]], i64* [[TMP39]], align 8
+// CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP40]], align 8
+// CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP42]], align 8
+// CHECK17-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP44]], align 8
+// CHECK17-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP45]], align 8
+// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126.region_id, i32 2, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK17-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK17:       omp_offload.failed11:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126(i64 [[TMP19]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK17:       omp_offload.cont12:
-// CHECK17-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 8
-// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 8
-// CHECK17-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP38]], align 8
-// CHECK17-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK17-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 2, i32* [[TMP49]], align 4
+// CHECK17-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP46]], i8*** [[TMP50]], align 8
+// CHECK17-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 8
+// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK17-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK17-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK17-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK17-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK17-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK17:       omp_offload.failed13:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126(i64 [[TMP35]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK17:       omp_offload.cont14:
+// CHECK17-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP59]], align 8
+// CHECK17-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP61]], align 8
+// CHECK17-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP62]], align 8
+// CHECK17-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP41:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l131.region_id, i32 1, i8** [[TMP39]], i8** [[TMP40]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP42:%.*]] = icmp ne i32 [[TMP41]], 0
-// CHECK17-NEXT:    br i1 [[TMP42]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK17:       omp_offload.failed17:
+// CHECK17-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP65]], align 4
+// CHECK17-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK17-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP63]], i8*** [[TMP67]], align 8
+// CHECK17-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 8
+// CHECK17-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP69]], align 8
+// CHECK17-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP70]], align 8
+// CHECK17-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP71]], align 8
+// CHECK17-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP72]], align 8
+// CHECK17-NEXT:    [[TMP73:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l131.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK17-NEXT:    [[TMP74:%.*]] = icmp ne i32 [[TMP73]], 0
+// CHECK17-NEXT:    br i1 [[TMP74]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK17:       omp_offload.failed20:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l131([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK17:       omp_offload.cont18:
-// CHECK17-NEXT:    [[TMP43:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    [[CONV20:%.*]] = bitcast i64* [[M_CASTED19]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP43]], i32* [[CONV20]], align 4
-// CHECK17-NEXT:    [[TMP44:%.*]] = load i64, i64* [[M_CASTED19]], align 8
-// CHECK17-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP44]], i64* [[TMP46]], align 8
-// CHECK17-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i64*
-// CHECK17-NEXT:    store i64 [[TMP44]], i64* [[TMP48]], align 8
-// CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP49]], align 8
-// CHECK17-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP51]], align 8
-// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP53]], align 8
-// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP54]], align 8
-// CHECK17-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK17:       omp_offload.cont21:
+// CHECK17-NEXT:    [[TMP75:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    [[CONV23:%.*]] = bitcast i64* [[M_CASTED22]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP75]], i32* [[CONV23]], align 4
+// CHECK17-NEXT:    [[TMP76:%.*]] = load i64, i64* [[M_CASTED22]], align 8
+// CHECK17-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP76]], i64* [[TMP78]], align 8
+// CHECK17-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
+// CHECK17-NEXT:    store i64 [[TMP76]], i64* [[TMP80]], align 8
+// CHECK17-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP81]], align 8
+// CHECK17-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP83]], align 8
+// CHECK17-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP85]], align 8
+// CHECK17-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP86]], align 8
+// CHECK17-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
-// CHECK17-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
-// CHECK17:       omp_offload.failed25:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136(i64 [[TMP44]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
-// CHECK17:       omp_offload.cont26:
+// CHECK17-NEXT:    [[KERNEL_ARGS28:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP89]], align 4
+// CHECK17-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 2, i32* [[TMP90]], align 4
+// CHECK17-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP87]], i8*** [[TMP91]], align 8
+// CHECK17-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP88]], i8*** [[TMP92]], align 8
+// CHECK17-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP93]], align 8
+// CHECK17-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP94]], align 8
+// CHECK17-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP95]], align 8
+// CHECK17-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP96]], align 8
+// CHECK17-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]])
+// CHECK17-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
+// CHECK17-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
+// CHECK17:       omp_offload.failed29:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136(i64 [[TMP76]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
+// CHECK17:       omp_offload.cont30:
 // CHECK17-NEXT:    ret i32 0
 //
 //
@@ -11120,31 +11970,31 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[M_CASTED:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[N_CASTED17:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS18:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES21:%.*]] = alloca [4 x i64], align 4
-// CHECK19-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_23:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
+// CHECK19-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[N_CASTED31:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS32:%.*]] = alloca [3 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS33:%.*]] = alloca [3 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS34:%.*]] = alloca [3 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES35:%.*]] = alloca [3 x i64], align 4
-// CHECK19-NEXT:    [[_TMP36:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_37:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_38:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[M_CASTED45:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[N_CASTED46:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS47:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS48:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS49:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES50:%.*]] = alloca [4 x i64], align 4
-// CHECK19-NEXT:    [[_TMP51:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_52:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_53:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[N_CASTED33:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS34:%.*]] = alloca [3 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS35:%.*]] = alloca [3 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS36:%.*]] = alloca [3 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES37:%.*]] = alloca [3 x i64], align 4
+// CHECK19-NEXT:    [[_TMP38:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_39:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_40:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[M_CASTED48:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[N_CASTED49:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS50:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS51:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS52:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES53:%.*]] = alloca [4 x i64], align 4
+// CHECK19-NEXT:    [[_TMP54:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_55:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_56:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK19-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK19-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -11202,258 +12052,343 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK19-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK19-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK19-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK19-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK19-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK19-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK19-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK19-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK19-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
-// CHECK19-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK19-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK19-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK19-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
-// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK19-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK19-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK19-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK19-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK19-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK19-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK19-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK19-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK19-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK19-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK19-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK19-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
+// CHECK19-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK19-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK19-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK19-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK19-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK19-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK19-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK19-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK19-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK19-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK19-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK19-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK19-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK19-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK19-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP59]])
-// CHECK19-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK19-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK19:       omp_offload.failed15:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK19:       omp_offload.cont16:
-// CHECK19-NEXT:    [[TMP62:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP62]], i32* [[M_CASTED]], align 4
-// CHECK19-NEXT:    [[TMP63:%.*]] = load i32, i32* [[M_CASTED]], align 4
-// CHECK19-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP64]], i32* [[N_CASTED17]], align 4
-// CHECK19-NEXT:    [[TMP65:%.*]] = load i32, i32* [[N_CASTED17]], align 4
-// CHECK19-NEXT:    [[TMP66:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP67:%.*]] = sext i32 [[TMP66]] to i64
-// CHECK19-NEXT:    [[TMP68:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES21]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP68]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
-// CHECK19-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP63]], i32* [[TMP70]], align 4
-// CHECK19-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP63]], i32* [[TMP72]], align 4
-// CHECK19-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP73]], align 4
-// CHECK19-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP65]], i32* [[TMP75]], align 4
-// CHECK19-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP65]], i32* [[TMP77]], align 4
-// CHECK19-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP78]], align 4
-// CHECK19-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP80]], align 4
-// CHECK19-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP82]], align 4
-// CHECK19-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP83]], align 4
-// CHECK19-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP85]], align 4
-// CHECK19-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP87]], align 4
-// CHECK19-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 3
-// CHECK19-NEXT:    store i64 [[TMP67]], i64* [[TMP88]], align 4
-// CHECK19-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK19-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK19-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP67]])
+// CHECK19-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK19-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK19-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK19-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK19-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK19-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK19-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK19-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK19-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK19-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK19-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK19:       omp_offload.failed16:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l153(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK19:       omp_offload.cont17:
+// CHECK19-NEXT:    [[TMP78:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP78]], i32* [[M_CASTED]], align 4
+// CHECK19-NEXT:    [[TMP79:%.*]] = load i32, i32* [[M_CASTED]], align 4
+// CHECK19-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP80]], i32* [[N_CASTED18]], align 4
+// CHECK19-NEXT:    [[TMP81:%.*]] = load i32, i32* [[N_CASTED18]], align 4
+// CHECK19-NEXT:    [[TMP82:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP83:%.*]] = sext i32 [[TMP82]] to i64
+// CHECK19-NEXT:    [[TMP84:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP84]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
+// CHECK19-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP79]], i32* [[TMP86]], align 4
+// CHECK19-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP88:%.*]] = bitcast i8** [[TMP87]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP79]], i32* [[TMP88]], align 4
+// CHECK19-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
 // CHECK19-NEXT:    store i8* null, i8** [[TMP89]], align 4
-// CHECK19-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP93:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP93]], i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK19-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK19-NEXT:    [[SUB25:%.*]] = sub nsw i32 [[TMP94]], 0
-// CHECK19-NEXT:    [[DIV26:%.*]] = sdiv i32 [[SUB25]], 1
-// CHECK19-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[DIV26]], 1
-// CHECK19-NEXT:    store i32 [[SUB27]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK19-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK19-NEXT:    [[ADD28:%.*]] = add nsw i32 [[TMP95]], 1
-// CHECK19-NEXT:    [[TMP96:%.*]] = zext i32 [[ADD28]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP96]])
-// CHECK19-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158.region_id, i32 4, i8** [[TMP90]], i8** [[TMP91]], i64* [[TMP92]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
-// CHECK19-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
-// CHECK19:       omp_offload.failed29:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158(i32 [[TMP63]], i32 [[TMP65]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
-// CHECK19:       omp_offload.cont30:
-// CHECK19-NEXT:    [[TMP99:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP99]], i32* [[N_CASTED31]], align 4
-// CHECK19-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N_CASTED31]], align 4
-// CHECK19-NEXT:    [[TMP101:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP102:%.*]] = sext i32 [[TMP101]] to i64
-// CHECK19-NEXT:    [[TMP103:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES35]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP103]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
-// CHECK19-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP100]], i32* [[TMP105]], align 4
-// CHECK19-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP100]], i32* [[TMP107]], align 4
-// CHECK19-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP108]], align 4
-// CHECK19-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP110]], align 4
-// CHECK19-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP112]], align 4
-// CHECK19-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP113]], align 4
-// CHECK19-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP115]], align 4
-// CHECK19-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP117]], align 4
-// CHECK19-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES35]], i32 0, i32 2
-// CHECK19-NEXT:    store i64 [[TMP102]], i64* [[TMP118]], align 4
-// CHECK19-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP119]], align 4
-// CHECK19-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES35]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP81]], i32* [[TMP91]], align 4
+// CHECK19-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP81]], i32* [[TMP93]], align 4
+// CHECK19-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP94]], align 4
+// CHECK19-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP96]], align 4
+// CHECK19-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP98]], align 4
+// CHECK19-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP99]], align 4
+// CHECK19-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP101]], align 4
+// CHECK19-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP103]], align 4
+// CHECK19-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 3
+// CHECK19-NEXT:    store i64 [[TMP83]], i64* [[TMP104]], align 4
+// CHECK19-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK19-NEXT:    store i8* null, i8** [[TMP105]], align 4
+// CHECK19-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP109:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP109]], i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK19-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK19-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP110]], 0
+// CHECK19-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
+// CHECK19-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
+// CHECK19-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK19-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK19-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP111]], 1
+// CHECK19-NEXT:    [[TMP112:%.*]] = zext i32 [[ADD29]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP112]])
+// CHECK19-NEXT:    [[KERNEL_ARGS30:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP113]], align 4
+// CHECK19-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 4, i32* [[TMP114]], align 4
+// CHECK19-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 4
+// CHECK19-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK19-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP108]], i64** [[TMP117]], align 4
+// CHECK19-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP118]], align 4
+// CHECK19-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP119]], align 4
+// CHECK19-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK19-NEXT:    [[TMP121:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]])
+// CHECK19-NEXT:    [[TMP122:%.*]] = icmp ne i32 [[TMP121]], 0
+// CHECK19-NEXT:    br i1 [[TMP122]], label [[OMP_OFFLOAD_FAILED31:%.*]], label [[OMP_OFFLOAD_CONT32:%.*]]
+// CHECK19:       omp_offload.failed31:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l158(i32 [[TMP79]], i32 [[TMP81]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT32]]
+// CHECK19:       omp_offload.cont32:
 // CHECK19-NEXT:    [[TMP123:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP123]], i32* [[DOTCAPTURE_EXPR_37]], align 4
-// CHECK19-NEXT:    [[TMP124:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_37]], align 4
-// CHECK19-NEXT:    [[SUB39:%.*]] = sub nsw i32 [[TMP124]], 0
-// CHECK19-NEXT:    [[DIV40:%.*]] = sdiv i32 [[SUB39]], 1
-// CHECK19-NEXT:    [[SUB41:%.*]] = sub nsw i32 [[DIV40]], 1
-// CHECK19-NEXT:    store i32 [[SUB41]], i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK19-NEXT:    [[TMP125:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK19-NEXT:    [[ADD42:%.*]] = add nsw i32 [[TMP125]], 1
-// CHECK19-NEXT:    [[TMP126:%.*]] = zext i32 [[ADD42]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP126]])
-// CHECK19-NEXT:    [[TMP127:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163.region_id, i32 3, i8** [[TMP120]], i8** [[TMP121]], i64* [[TMP122]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP128:%.*]] = icmp ne i32 [[TMP127]], 0
-// CHECK19-NEXT:    br i1 [[TMP128]], label [[OMP_OFFLOAD_FAILED43:%.*]], label [[OMP_OFFLOAD_CONT44:%.*]]
-// CHECK19:       omp_offload.failed43:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163(i32 [[TMP100]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT44]]
-// CHECK19:       omp_offload.cont44:
-// CHECK19-NEXT:    [[TMP129:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP129]], i32* [[M_CASTED45]], align 4
-// CHECK19-NEXT:    [[TMP130:%.*]] = load i32, i32* [[M_CASTED45]], align 4
-// CHECK19-NEXT:    [[TMP131:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP131]], i32* [[N_CASTED46]], align 4
-// CHECK19-NEXT:    [[TMP132:%.*]] = load i32, i32* [[N_CASTED46]], align 4
-// CHECK19-NEXT:    [[TMP133:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP134:%.*]] = sext i32 [[TMP133]] to i64
-// CHECK19-NEXT:    [[TMP135:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES50]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP135]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
-// CHECK19-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP130]], i32* [[TMP137]], align 4
-// CHECK19-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP130]], i32* [[TMP139]], align 4
-// CHECK19-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP140]], align 4
-// CHECK19-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP132]], i32* [[TMP142]], align 4
-// CHECK19-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP132]], i32* [[TMP144]], align 4
-// CHECK19-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP145]], align 4
-// CHECK19-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP147]], align 4
-// CHECK19-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP149]], align 4
-// CHECK19-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP150]], align 4
-// CHECK19-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP152]], align 4
-// CHECK19-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP154]], align 4
-// CHECK19-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES50]], i32 0, i32 3
-// CHECK19-NEXT:    store i64 [[TMP134]], i64* [[TMP155]], align 4
-// CHECK19-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 3
-// CHECK19-NEXT:    store i8* null, i8** [[TMP156]], align 4
-// CHECK19-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES50]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP160:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP160]], i32* [[DOTCAPTURE_EXPR_52]], align 4
-// CHECK19-NEXT:    [[TMP161:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_52]], align 4
-// CHECK19-NEXT:    [[SUB54:%.*]] = sub nsw i32 [[TMP161]], 0
-// CHECK19-NEXT:    [[DIV55:%.*]] = sdiv i32 [[SUB54]], 1
-// CHECK19-NEXT:    [[SUB56:%.*]] = sub nsw i32 [[DIV55]], 1
-// CHECK19-NEXT:    store i32 [[SUB56]], i32* [[DOTCAPTURE_EXPR_53]], align 4
-// CHECK19-NEXT:    [[TMP162:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_53]], align 4
-// CHECK19-NEXT:    [[ADD57:%.*]] = add nsw i32 [[TMP162]], 1
-// CHECK19-NEXT:    [[TMP163:%.*]] = zext i32 [[ADD57]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP163]])
-// CHECK19-NEXT:    [[TMP164:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168.region_id, i32 4, i8** [[TMP157]], i8** [[TMP158]], i64* [[TMP159]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP165:%.*]] = icmp ne i32 [[TMP164]], 0
-// CHECK19-NEXT:    br i1 [[TMP165]], label [[OMP_OFFLOAD_FAILED58:%.*]], label [[OMP_OFFLOAD_CONT59:%.*]]
-// CHECK19:       omp_offload.failed58:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168(i32 [[TMP130]], i32 [[TMP132]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT59]]
-// CHECK19:       omp_offload.cont59:
-// CHECK19-NEXT:    [[TMP166:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK19-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP166]])
+// CHECK19-NEXT:    store i32 [[TMP123]], i32* [[N_CASTED33]], align 4
+// CHECK19-NEXT:    [[TMP124:%.*]] = load i32, i32* [[N_CASTED33]], align 4
+// CHECK19-NEXT:    [[TMP125:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP126:%.*]] = sext i32 [[TMP125]] to i64
+// CHECK19-NEXT:    [[TMP127:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES37]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP127]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
+// CHECK19-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP124]], i32* [[TMP129]], align 4
+// CHECK19-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP124]], i32* [[TMP131]], align 4
+// CHECK19-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP132]], align 4
+// CHECK19-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP134]], align 4
+// CHECK19-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP136]], align 4
+// CHECK19-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP137]], align 4
+// CHECK19-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP139]], align 4
+// CHECK19-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP141]], align 4
+// CHECK19-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES37]], i32 0, i32 2
+// CHECK19-NEXT:    store i64 [[TMP126]], i64* [[TMP142]], align 4
+// CHECK19-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP143]], align 4
+// CHECK19-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES37]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP147:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP147]], i32* [[DOTCAPTURE_EXPR_39]], align 4
+// CHECK19-NEXT:    [[TMP148:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_39]], align 4
+// CHECK19-NEXT:    [[SUB41:%.*]] = sub nsw i32 [[TMP148]], 0
+// CHECK19-NEXT:    [[DIV42:%.*]] = sdiv i32 [[SUB41]], 1
+// CHECK19-NEXT:    [[SUB43:%.*]] = sub nsw i32 [[DIV42]], 1
+// CHECK19-NEXT:    store i32 [[SUB43]], i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK19-NEXT:    [[TMP149:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK19-NEXT:    [[ADD44:%.*]] = add nsw i32 [[TMP149]], 1
+// CHECK19-NEXT:    [[TMP150:%.*]] = zext i32 [[ADD44]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP150]])
+// CHECK19-NEXT:    [[KERNEL_ARGS45:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP151]], align 4
+// CHECK19-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP152]], align 4
+// CHECK19-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP144]], i8*** [[TMP153]], align 4
+// CHECK19-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP145]], i8*** [[TMP154]], align 4
+// CHECK19-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP146]], i64** [[TMP155]], align 4
+// CHECK19-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP156]], align 4
+// CHECK19-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP157]], align 4
+// CHECK19-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP158]], align 4
+// CHECK19-NEXT:    [[TMP159:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]])
+// CHECK19-NEXT:    [[TMP160:%.*]] = icmp ne i32 [[TMP159]], 0
+// CHECK19-NEXT:    br i1 [[TMP160]], label [[OMP_OFFLOAD_FAILED46:%.*]], label [[OMP_OFFLOAD_CONT47:%.*]]
+// CHECK19:       omp_offload.failed46:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l163(i32 [[TMP124]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT47]]
+// CHECK19:       omp_offload.cont47:
+// CHECK19-NEXT:    [[TMP161:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP161]], i32* [[M_CASTED48]], align 4
+// CHECK19-NEXT:    [[TMP162:%.*]] = load i32, i32* [[M_CASTED48]], align 4
+// CHECK19-NEXT:    [[TMP163:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP163]], i32* [[N_CASTED49]], align 4
+// CHECK19-NEXT:    [[TMP164:%.*]] = load i32, i32* [[N_CASTED49]], align 4
+// CHECK19-NEXT:    [[TMP165:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP166:%.*]] = sext i32 [[TMP165]] to i64
+// CHECK19-NEXT:    [[TMP167:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES53]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP167]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
+// CHECK19-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP162]], i32* [[TMP169]], align 4
+// CHECK19-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP171:%.*]] = bitcast i8** [[TMP170]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP162]], i32* [[TMP171]], align 4
+// CHECK19-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP172]], align 4
+// CHECK19-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP174:%.*]] = bitcast i8** [[TMP173]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP164]], i32* [[TMP174]], align 4
+// CHECK19-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP176:%.*]] = bitcast i8** [[TMP175]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP164]], i32* [[TMP176]], align 4
+// CHECK19-NEXT:    [[TMP177:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP177]], align 4
+// CHECK19-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP179]], align 4
+// CHECK19-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP181]], align 4
+// CHECK19-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP182]], align 4
+// CHECK19-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP184:%.*]] = bitcast i8** [[TMP183]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP184]], align 4
+// CHECK19-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP186:%.*]] = bitcast i8** [[TMP185]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP186]], align 4
+// CHECK19-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES53]], i32 0, i32 3
+// CHECK19-NEXT:    store i64 [[TMP166]], i64* [[TMP187]], align 4
+// CHECK19-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 3
+// CHECK19-NEXT:    store i8* null, i8** [[TMP188]], align 4
+// CHECK19-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES53]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP192:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP192]], i32* [[DOTCAPTURE_EXPR_55]], align 4
+// CHECK19-NEXT:    [[TMP193:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_55]], align 4
+// CHECK19-NEXT:    [[SUB57:%.*]] = sub nsw i32 [[TMP193]], 0
+// CHECK19-NEXT:    [[DIV58:%.*]] = sdiv i32 [[SUB57]], 1
+// CHECK19-NEXT:    [[SUB59:%.*]] = sub nsw i32 [[DIV58]], 1
+// CHECK19-NEXT:    store i32 [[SUB59]], i32* [[DOTCAPTURE_EXPR_56]], align 4
+// CHECK19-NEXT:    [[TMP194:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_56]], align 4
+// CHECK19-NEXT:    [[ADD60:%.*]] = add nsw i32 [[TMP194]], 1
+// CHECK19-NEXT:    [[TMP195:%.*]] = zext i32 [[ADD60]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP195]])
+// CHECK19-NEXT:    [[KERNEL_ARGS61:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP196]], align 4
+// CHECK19-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 4, i32* [[TMP197]], align 4
+// CHECK19-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP189]], i8*** [[TMP198]], align 4
+// CHECK19-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP190]], i8*** [[TMP199]], align 4
+// CHECK19-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP191]], i64** [[TMP200]], align 4
+// CHECK19-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP201]], align 4
+// CHECK19-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP202]], align 4
+// CHECK19-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP203]], align 4
+// CHECK19-NEXT:    [[TMP204:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]])
+// CHECK19-NEXT:    [[TMP205:%.*]] = icmp ne i32 [[TMP204]], 0
+// CHECK19-NEXT:    br i1 [[TMP205]], label [[OMP_OFFLOAD_FAILED62:%.*]], label [[OMP_OFFLOAD_CONT63:%.*]]
+// CHECK19:       omp_offload.failed62:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l168(i32 [[TMP162]], i32 [[TMP164]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT63]]
+// CHECK19:       omp_offload.cont63:
+// CHECK19-NEXT:    [[TMP206:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK19-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP206]])
 // CHECK19-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK19-NEXT:    [[TMP167:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP167]])
-// CHECK19-NEXT:    [[TMP168:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK19-NEXT:    ret i32 [[TMP168]]
+// CHECK19-NEXT:    [[TMP207:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP207]])
+// CHECK19-NEXT:    [[TMP208:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK19-NEXT:    ret i32 [[TMP208]]
 //
 //
 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l148
@@ -12553,19 +13488,19 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 4
 // CHECK19-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[M_CASTED:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 4
-// CHECK19-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[M_CASTED19:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 4
+// CHECK19-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[M_CASTED22:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK19-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK19-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -12579,107 +13514,192 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK19-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK19-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK19-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK19-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK19-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK19-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK19-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK19-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l116([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
-// CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK19-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK19-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK19-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l121.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK19-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK19:       omp_offload.failed5:
+// CHECK19-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK19-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK19-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l121.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK19-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK19-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK19:       omp_offload.failed6:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l121([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK19:       omp_offload.cont6:
-// CHECK19-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP18]], i32* [[M_CASTED]], align 4
-// CHECK19-NEXT:    [[TMP19:%.*]] = load i32, i32* [[M_CASTED]], align 4
-// CHECK19-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP19]], i32* [[TMP21]], align 4
-// CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP19]], i32* [[TMP23]], align 4
-// CHECK19-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP24]], align 4
-// CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP26]], align 4
-// CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP28]], align 4
-// CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP29]], align 4
-// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK19:       omp_offload.cont7:
+// CHECK19-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP34]], i32* [[M_CASTED]], align 4
+// CHECK19-NEXT:    [[TMP35:%.*]] = load i32, i32* [[M_CASTED]], align 4
+// CHECK19-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP35]], i32* [[TMP37]], align 4
+// CHECK19-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP35]], i32* [[TMP39]], align 4
+// CHECK19-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP40]], align 4
+// CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP42]], align 4
+// CHECK19-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP44]], align 4
+// CHECK19-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP45]], align 4
+// CHECK19-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126.region_id, i32 2, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK19-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK19:       omp_offload.failed11:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126(i32 [[TMP19]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK19:       omp_offload.cont12:
-// CHECK19-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 4
-// CHECK19-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 4
-// CHECK19-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP38]], align 4
-// CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK19-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 2, i32* [[TMP49]], align 4
+// CHECK19-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP46]], i8*** [[TMP50]], align 4
+// CHECK19-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 4
+// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP52]], align 4
+// CHECK19-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP53]], align 4
+// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP54]], align 4
+// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP55]], align 4
+// CHECK19-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK19-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK19-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK19:       omp_offload.failed13:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l126(i32 [[TMP35]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK19:       omp_offload.cont14:
+// CHECK19-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP59]], align 4
+// CHECK19-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP61]], align 4
+// CHECK19-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP62]], align 4
+// CHECK19-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP41:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l131.region_id, i32 1, i8** [[TMP39]], i8** [[TMP40]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP42:%.*]] = icmp ne i32 [[TMP41]], 0
-// CHECK19-NEXT:    br i1 [[TMP42]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK19:       omp_offload.failed17:
+// CHECK19-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP65]], align 4
+// CHECK19-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK19-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP63]], i8*** [[TMP67]], align 4
+// CHECK19-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK19-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP69]], align 4
+// CHECK19-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK19-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP71]], align 4
+// CHECK19-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK19-NEXT:    [[TMP73:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l131.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK19-NEXT:    [[TMP74:%.*]] = icmp ne i32 [[TMP73]], 0
+// CHECK19-NEXT:    br i1 [[TMP74]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK19:       omp_offload.failed20:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l131([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK19:       omp_offload.cont18:
-// CHECK19-NEXT:    [[TMP43:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP43]], i32* [[M_CASTED19]], align 4
-// CHECK19-NEXT:    [[TMP44:%.*]] = load i32, i32* [[M_CASTED19]], align 4
-// CHECK19-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP44]], i32* [[TMP46]], align 4
-// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP44]], i32* [[TMP48]], align 4
-// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP49]], align 4
-// CHECK19-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP51]], align 4
-// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP53]], align 4
-// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP54]], align 4
-// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK19:       omp_offload.cont21:
+// CHECK19-NEXT:    [[TMP75:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP75]], i32* [[M_CASTED22]], align 4
+// CHECK19-NEXT:    [[TMP76:%.*]] = load i32, i32* [[M_CASTED22]], align 4
+// CHECK19-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP76]], i32* [[TMP78]], align 4
+// CHECK19-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP76]], i32* [[TMP80]], align 4
+// CHECK19-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP81]], align 4
+// CHECK19-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP83]], align 4
+// CHECK19-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP85]], align 4
+// CHECK19-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP86]], align 4
+// CHECK19-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
-// CHECK19-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
-// CHECK19:       omp_offload.failed24:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136(i32 [[TMP44]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
-// CHECK19:       omp_offload.cont25:
+// CHECK19-NEXT:    [[KERNEL_ARGS27:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP89]], align 4
+// CHECK19-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 2, i32* [[TMP90]], align 4
+// CHECK19-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP87]], i8*** [[TMP91]], align 4
+// CHECK19-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP88]], i8*** [[TMP92]], align 4
+// CHECK19-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP93]], align 4
+// CHECK19-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP94]], align 4
+// CHECK19-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP95]], align 4
+// CHECK19-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP96]], align 4
+// CHECK19-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]])
+// CHECK19-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
+// CHECK19-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
+// CHECK19:       omp_offload.failed28:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l136(i32 [[TMP76]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
+// CHECK19:       omp_offload.cont29:
 // CHECK19-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_simd_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_simd_codegen.cpp
index c3e025f..4e95ccc 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_simd_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_simd_codegen.cpp
@@ -268,67 +268,101 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP32]], 1
 // CHECK1-NEXT:    [[TMP33:%.*]] = zext i32 [[ADD]] to i64
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP33]])
-// CHECK1-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP29]], i32 0)
-// CHECK1-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK1-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP35]], align 4
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP40]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP41]], align 8
+// CHECK1-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 [[TMP29]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK1-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29(i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], [100 x i32]* @a) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP36:%.*]] = load i32, i32* [[I]], align 4
+// CHECK1-NEXT:    [[TMP44:%.*]] = load i32, i32* [[I]], align 4
 // CHECK1-NEXT:    [[CONV6:%.*]] = bitcast i64* [[I_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP36]], i32* [[CONV6]], align 4
-// CHECK1-NEXT:    [[TMP37:%.*]] = load i64, i64* [[I_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP38:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    store i32 [[TMP44]], i32* [[CONV6]], align 4
+// CHECK1-NEXT:    [[TMP45:%.*]] = load i64, i64* [[I_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = load i32, i32* [[N_ADDR]], align 4
 // CHECK1-NEXT:    [[CONV8:%.*]] = bitcast i64* [[N_CASTED7]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP38]], i32* [[CONV8]], align 4
-// CHECK1-NEXT:    [[TMP39:%.*]] = load i64, i64* [[N_CASTED7]], align 8
-// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to [100 x i32]**
-// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP41]], align 8
-// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to [100 x i32]**
-// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP43]], align 8
-// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP44]], align 8
-// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP37]], i64* [[TMP46]], align 8
-// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP37]], i64* [[TMP48]], align 8
-// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP49]], align 8
-// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP39]], i64* [[TMP51]], align 8
-// CHECK1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 2
-// CHECK1-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP39]], i64* [[TMP53]], align 8
-// CHECK1-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 2
-// CHECK1-NEXT:    store i8* null, i8** [[TMP54]], align 8
-// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP57:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    store i32 [[TMP57]], i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK1-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK1-NEXT:    [[SUB15:%.*]] = sub nsw i32 [[TMP58]], 0
+// CHECK1-NEXT:    store i32 [[TMP46]], i32* [[CONV8]], align 4
+// CHECK1-NEXT:    [[TMP47:%.*]] = load i64, i64* [[N_CASTED7]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to [100 x i32]**
+// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP49]], align 8
+// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to [100 x i32]**
+// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP51]], align 8
+// CHECK1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP52]], align 8
+// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP45]], i64* [[TMP54]], align 8
+// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP45]], i64* [[TMP56]], align 8
+// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP57]], align 8
+// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP47]], i64* [[TMP59]], align 8
+// CHECK1-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 2
+// CHECK1-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP47]], i64* [[TMP61]], align 8
+// CHECK1-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 2
+// CHECK1-NEXT:    store i8* null, i8** [[TMP62]], align 8
+// CHECK1-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP65:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    store i32 [[TMP65]], i32* [[DOTCAPTURE_EXPR_13]], align 4
+// CHECK1-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
+// CHECK1-NEXT:    [[SUB15:%.*]] = sub nsw i32 [[TMP66]], 0
 // CHECK1-NEXT:    [[DIV16:%.*]] = sdiv i32 [[SUB15]], 1
 // CHECK1-NEXT:    [[SUB17:%.*]] = sub nsw i32 [[DIV16]], 1
 // CHECK1-NEXT:    store i32 [[SUB17]], i32* [[DOTCAPTURE_EXPR_14]], align 4
-// CHECK1-NEXT:    [[TMP59:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_14]], align 4
-// CHECK1-NEXT:    [[ADD18:%.*]] = add nsw i32 [[TMP59]], 1
-// CHECK1-NEXT:    [[TMP60:%.*]] = zext i32 [[ADD18]] to i64
-// CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP60]])
-// CHECK1-NEXT:    [[TMP61:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36.region_id, i32 3, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP62:%.*]] = icmp ne i32 [[TMP61]], 0
-// CHECK1-NEXT:    br i1 [[TMP62]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
-// CHECK1:       omp_offload.failed19:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36([100 x i32]* @a, i64 [[TMP37]], i64 [[TMP39]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
-// CHECK1:       omp_offload.cont20:
-// CHECK1-NEXT:    [[TMP63:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i64 0, i64 0), align 4
-// CHECK1-NEXT:    ret i32 [[TMP63]]
+// CHECK1-NEXT:    [[TMP67:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_14]], align 4
+// CHECK1-NEXT:    [[ADD18:%.*]] = add nsw i32 [[TMP67]], 1
+// CHECK1-NEXT:    [[TMP68:%.*]] = zext i32 [[ADD18]] to i64
+// CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP68]])
+// CHECK1-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP69]], align 4
+// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 3, i32* [[TMP70]], align 4
+// CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP63]], i8*** [[TMP71]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP64]], i8*** [[TMP72]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP73]], align 8
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP74]], align 8
+// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP75]], align 8
+// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP76]], align 8
+// CHECK1-NEXT:    [[TMP77:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK1-NEXT:    [[TMP78:%.*]] = icmp ne i32 [[TMP77]], 0
+// CHECK1-NEXT:    br i1 [[TMP78]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK1:       omp_offload.failed20:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36([100 x i32]* @a, i64 [[TMP45]], i64 [[TMP47]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK1:       omp_offload.cont21:
+// CHECK1-NEXT:    [[TMP79:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i64 0, i64 0), align 4
+// CHECK1-NEXT:    ret i32 [[TMP79]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29
@@ -912,65 +946,99 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP32]], 1
 // CHECK3-NEXT:    [[TMP33:%.*]] = zext i32 [[ADD]] to i64
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP33]])
-// CHECK3-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP29]], i32 0)
-// CHECK3-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK3-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP39]], align 4
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP40]], align 4
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 [[TMP29]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP43:%.*]] = icmp ne i32 [[TMP42]], 0
+// CHECK3-NEXT:    br i1 [[TMP43]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29(i32 [[TMP2]], i32 [[TMP4]], i32 [[TMP6]], [100 x i32]* @a) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP36:%.*]] = load i32, i32* [[I]], align 4
-// CHECK3-NEXT:    store i32 [[TMP36]], i32* [[I_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP37:%.*]] = load i32, i32* [[I_CASTED]], align 4
-// CHECK3-NEXT:    [[TMP38:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    store i32 [[TMP38]], i32* [[N_CASTED4]], align 4
-// CHECK3-NEXT:    [[TMP39:%.*]] = load i32, i32* [[N_CASTED4]], align 4
-// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP41:%.*]] = bitcast i8** [[TMP40]] to [100 x i32]**
-// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP41]], align 4
-// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to [100 x i32]**
-// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP43]], align 4
-// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP44]], align 4
-// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP37]], i32* [[TMP46]], align 4
-// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP37]], i32* [[TMP48]], align 4
-// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP49]], align 4
-// CHECK3-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP39]], i32* [[TMP51]], align 4
-// CHECK3-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK3-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP39]], i32* [[TMP53]], align 4
-// CHECK3-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 2
-// CHECK3-NEXT:    store i8* null, i8** [[TMP54]], align 4
-// CHECK3-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP57:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    store i32 [[TMP57]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK3-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK3-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP58]], 0
+// CHECK3-NEXT:    [[TMP44:%.*]] = load i32, i32* [[I]], align 4
+// CHECK3-NEXT:    store i32 [[TMP44]], i32* [[I_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = load i32, i32* [[I_CASTED]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    store i32 [[TMP46]], i32* [[N_CASTED4]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = load i32, i32* [[N_CASTED4]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to [100 x i32]**
+// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP49]], align 4
+// CHECK3-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to [100 x i32]**
+// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP51]], align 4
+// CHECK3-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP52]], align 4
+// CHECK3-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP45]], i32* [[TMP54]], align 4
+// CHECK3-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP45]], i32* [[TMP56]], align 4
+// CHECK3-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP57]], align 4
+// CHECK3-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP47]], i32* [[TMP59]], align 4
+// CHECK3-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK3-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP47]], i32* [[TMP61]], align 4
+// CHECK3-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8* null, i8** [[TMP62]], align 4
+// CHECK3-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP65:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    store i32 [[TMP65]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK3-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK3-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP66]], 0
 // CHECK3-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK3-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK3-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK3-NEXT:    [[TMP59:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK3-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP59]], 1
-// CHECK3-NEXT:    [[TMP60:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP60]])
-// CHECK3-NEXT:    [[TMP61:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36.region_id, i32 3, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP62:%.*]] = icmp ne i32 [[TMP61]], 0
-// CHECK3-NEXT:    br i1 [[TMP62]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK3:       omp_offload.failed15:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36([100 x i32]* @a, i32 [[TMP37]], i32 [[TMP39]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK3:       omp_offload.cont16:
-// CHECK3-NEXT:    [[TMP63:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i32 0, i32 0), align 4
-// CHECK3-NEXT:    ret i32 [[TMP63]]
+// CHECK3-NEXT:    [[TMP67:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK3-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP67]], 1
+// CHECK3-NEXT:    [[TMP68:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP68]])
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 3, i32* [[TMP70]], align 4
+// CHECK3-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP63]], i8*** [[TMP71]], align 4
+// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP64]], i8*** [[TMP72]], align 4
+// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK3-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP74]], align 4
+// CHECK3-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK3-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP76]], align 4
+// CHECK3-NEXT:    [[TMP77:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP78:%.*]] = icmp ne i32 [[TMP77]], 0
+// CHECK3-NEXT:    br i1 [[TMP78]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36([100 x i32]* @a, i32 [[TMP45]], i32 [[TMP47]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[TMP79:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i32 0, i32 0), align 4
+// CHECK3-NEXT:    ret i32 [[TMP79]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l29
@@ -1805,18 +1873,35 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP35]], 1
 // CHECK9-NEXT:    [[TMP36:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 [[TMP36]])
-// CHECK9-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l74.region_id, i32 4, i8** [[TMP30]], i8** [[TMP31]], i64* [[TMP32]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
-// CHECK9-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP37]], align 4
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP38]], align 4
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP30]], i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP31]], i8*** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP32]], i64** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP44]], align 8
+// CHECK9-NEXT:    [[TMP45:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l74.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP46:%.*]] = icmp ne i32 [[TMP45]], 0
+// CHECK9-NEXT:    br i1 [[TMP46]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l74(i64 [[TMP1]], i32* [[VLA]], i64 [[TMP4]], i64 [[TMP6]]) #[[ATTR5:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK9-NEXT:    [[TMP40:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP40]])
-// CHECK9-NEXT:    ret i32 [[TMP39]]
+// CHECK9-NEXT:    [[TMP47:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK9-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
+// CHECK9-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l74
@@ -2166,18 +2251,35 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP35]], 1
 // CHECK11-NEXT:    [[TMP36:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 [[TMP36]])
-// CHECK11-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l74.region_id, i32 4, i8** [[TMP30]], i8** [[TMP31]], i64* [[TMP32]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
-// CHECK11-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP30]], i8*** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP31]], i8*** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP32]], i64** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP44]], align 4
+// CHECK11-NEXT:    [[TMP45:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l74.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP46:%.*]] = icmp ne i32 [[TMP45]], 0
+// CHECK11-NEXT:    br i1 [[TMP46]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l74(i32 [[TMP0]], i32* [[VLA]], i32 [[TMP3]], i32 [[TMP5]]) #[[ATTR5:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i32 0
-// CHECK11-NEXT:    [[TMP39:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK11-NEXT:    [[TMP40:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP40]])
-// CHECK11-NEXT:    ret i32 [[TMP39]]
+// CHECK11-NEXT:    [[TMP47:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
+// CHECK11-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l74
@@ -2648,17 +2750,34 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 123)
-// CHECK17-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l112.region_id, i32 2, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK17-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP14]], align 4
+// CHECK17-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 2, i32* [[TMP15]], align 4
+// CHECK17-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP12]], i8*** [[TMP16]], align 8
+// CHECK17-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP13]], i8*** [[TMP17]], align 8
+// CHECK17-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP18]], align 8
+// CHECK17-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP19]], align 8
+// CHECK17-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP20]], align 8
+// CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP21]], align 8
+// CHECK17-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l112.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
+// CHECK17-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l112(%struct.SS* [[THIS1]], i64 [[TMP1]]) #[[ATTR4:[0-9]+]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
 // CHECK17-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK17-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A2]], i64 0, i64 0
-// CHECK17-NEXT:    [[TMP16:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK17-NEXT:    ret i32 [[TMP16]]
+// CHECK17-NEXT:    [[TMP24:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK17-NEXT:    ret i32 [[TMP24]]
 //
 //
 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l112
@@ -2902,17 +3021,34 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 123)
-// CHECK19-NEXT:    [[TMP14:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l112.region_id, i32 2, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP15:%.*]] = icmp ne i32 [[TMP14]], 0
-// CHECK19-NEXT:    br i1 [[TMP15]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP14]], align 4
+// CHECK19-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 2, i32* [[TMP15]], align 4
+// CHECK19-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP12]], i8*** [[TMP16]], align 4
+// CHECK19-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP13]], i8*** [[TMP17]], align 4
+// CHECK19-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP18]], align 4
+// CHECK19-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP19]], align 4
+// CHECK19-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP20]], align 4
+// CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP21]], align 4
+// CHECK19-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l112.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
+// CHECK19-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l112(%struct.SS* [[THIS1]], i32 [[TMP1]]) #[[ATTR4:[0-9]+]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
 // CHECK19-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK19-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A2]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP16:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK19-NEXT:    ret i32 [[TMP16]]
+// CHECK19-NEXT:    [[TMP24:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK19-NEXT:    ret i32 [[TMP24]]
 //
 //
 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l112
@@ -3320,20 +3456,37 @@ int main (int argc, char **argv) {
 // CHECK25-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP35]], 1
 // CHECK25-NEXT:    [[TMP36:%.*]] = zext i32 [[ADD]] to i64
 // CHECK25-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 [[TMP36]])
-// CHECK25-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l166.region_id, i32 4, i8** [[TMP30]], i8** [[TMP31]], i64* [[TMP32]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK25-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
-// CHECK25-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK25-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK25-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK25-NEXT:    store i32 1, i32* [[TMP37]], align 4
+// CHECK25-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK25-NEXT:    store i32 4, i32* [[TMP38]], align 4
+// CHECK25-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK25-NEXT:    store i8** [[TMP30]], i8*** [[TMP39]], align 8
+// CHECK25-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK25-NEXT:    store i8** [[TMP31]], i8*** [[TMP40]], align 8
+// CHECK25-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK25-NEXT:    store i64* [[TMP32]], i64** [[TMP41]], align 8
+// CHECK25-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK25-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP42]], align 8
+// CHECK25-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK25-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK25-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK25-NEXT:    store i8** null, i8*** [[TMP44]], align 8
+// CHECK25-NEXT:    [[TMP45:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l166.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK25-NEXT:    [[TMP46:%.*]] = icmp ne i32 [[TMP45]], 0
+// CHECK25-NEXT:    br i1 [[TMP46]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK25:       omp_offload.failed:
 // CHECK25-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l166(i64 [[TMP1]], i32* [[VLA]], i64 [[TMP4]], i64 [[TMP6]]) #[[ATTR5:[0-9]+]]
 // CHECK25-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK25:       omp_offload.cont:
-// CHECK25-NEXT:    [[TMP39:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK25-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP39]])
+// CHECK25-NEXT:    [[TMP47:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK25-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP47]])
 // CHECK25-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK25-NEXT:    [[TMP40:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK25-NEXT:    call void @llvm.stackrestore(i8* [[TMP40]])
-// CHECK25-NEXT:    [[TMP41:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK25-NEXT:    ret i32 [[TMP41]]
+// CHECK25-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK25-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
+// CHECK25-NEXT:    [[TMP49:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK25-NEXT:    ret i32 [[TMP49]]
 //
 //
 // CHECK25-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l166
@@ -3648,9 +3801,26 @@ int main (int argc, char **argv) {
 // CHECK25-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK25-NEXT:    [[TMP21:%.*]] = load i32, i32* [[TE]], align 4
 // CHECK25-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 10)
-// CHECK25-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l155.region_id, i32 3, i8** [[TMP19]], i8** [[TMP20]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 [[TMP21]], i32 0)
-// CHECK25-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK25-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK25-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK25-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK25-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK25-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK25-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK25-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK25-NEXT:    store i8** [[TMP19]], i8*** [[TMP24]], align 8
+// CHECK25-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK25-NEXT:    store i8** [[TMP20]], i8*** [[TMP25]], align 8
+// CHECK25-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK25-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK25-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK25-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK25-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK25-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK25-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK25-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK25-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 [[TMP21]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l155.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK25-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK25-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK25:       omp_offload.failed:
 // CHECK25-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l155(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[A]]) #[[ATTR5]]
 // CHECK25-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3925,20 +4095,37 @@ int main (int argc, char **argv) {
 // CHECK27-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP35]], 1
 // CHECK27-NEXT:    [[TMP36:%.*]] = zext i32 [[ADD]] to i64
 // CHECK27-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 [[TMP36]])
-// CHECK27-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l166.region_id, i32 4, i8** [[TMP30]], i8** [[TMP31]], i64* [[TMP32]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK27-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
-// CHECK27-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK27-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK27-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK27-NEXT:    store i32 1, i32* [[TMP37]], align 4
+// CHECK27-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK27-NEXT:    store i32 4, i32* [[TMP38]], align 4
+// CHECK27-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK27-NEXT:    store i8** [[TMP30]], i8*** [[TMP39]], align 4
+// CHECK27-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK27-NEXT:    store i8** [[TMP31]], i8*** [[TMP40]], align 4
+// CHECK27-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK27-NEXT:    store i64* [[TMP32]], i64** [[TMP41]], align 4
+// CHECK27-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK27-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP42]], align 4
+// CHECK27-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK27-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK27-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK27-NEXT:    store i8** null, i8*** [[TMP44]], align 4
+// CHECK27-NEXT:    [[TMP45:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l166.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK27-NEXT:    [[TMP46:%.*]] = icmp ne i32 [[TMP45]], 0
+// CHECK27-NEXT:    br i1 [[TMP46]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK27:       omp_offload.failed:
 // CHECK27-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l166(i32 [[TMP0]], i32* [[VLA]], i32 [[TMP3]], i32 [[TMP5]]) #[[ATTR5:[0-9]+]]
 // CHECK27-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK27:       omp_offload.cont:
-// CHECK27-NEXT:    [[TMP39:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK27-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP39]])
+// CHECK27-NEXT:    [[TMP47:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK27-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP47]])
 // CHECK27-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK27-NEXT:    [[TMP40:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK27-NEXT:    call void @llvm.stackrestore(i8* [[TMP40]])
-// CHECK27-NEXT:    [[TMP41:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK27-NEXT:    ret i32 [[TMP41]]
+// CHECK27-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK27-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
+// CHECK27-NEXT:    [[TMP49:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK27-NEXT:    ret i32 [[TMP49]]
 //
 //
 // CHECK27-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l166
@@ -4244,9 +4431,26 @@ int main (int argc, char **argv) {
 // CHECK27-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK27-NEXT:    [[TMP21:%.*]] = load i32, i32* [[TE]], align 4
 // CHECK27-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 10)
-// CHECK27-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l155.region_id, i32 3, i8** [[TMP19]], i8** [[TMP20]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 [[TMP21]], i32 0)
-// CHECK27-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK27-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK27-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK27-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK27-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK27-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK27-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK27-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK27-NEXT:    store i8** [[TMP19]], i8*** [[TMP24]], align 4
+// CHECK27-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK27-NEXT:    store i8** [[TMP20]], i8*** [[TMP25]], align 4
+// CHECK27-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK27-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK27-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK27-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK27-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK27-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK27-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK27-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK27-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 [[TMP21]], i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l155.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK27-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK27-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK27:       omp_offload.failed:
 // CHECK27-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l155(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[A]]) #[[ATTR5]]
 // CHECK27-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_simd_collapse_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_simd_collapse_codegen.cpp
index 97b36d9..7267e0d 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_simd_collapse_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_simd_collapse_codegen.cpp
@@ -138,9 +138,26 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 56088)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -148,8 +165,8 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i64 0, i64 0
 // CHECK1-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP9]]
+// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -373,9 +390,26 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 56088)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -383,8 +417,8 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP9]]
+// CHECK3-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -823,20 +857,37 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP46:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_4]], align 8
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP46]], 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK9-NEXT:    [[TMP47:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l86.region_id, i32 5, i8** [[TMP39]], i8** [[TMP40]], i64* [[TMP41]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP48:%.*]] = icmp ne i32 [[TMP47]], 0
-// CHECK9-NEXT:    br i1 [[TMP48]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP47]], align 4
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP48]], align 4
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 8
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP40]], i8*** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP41]], i64** [[TMP51]], align 8
+// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l86.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP56:%.*]] = icmp ne i32 [[TMP55]], 0
+// CHECK9-NEXT:    br i1 [[TMP56]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l86(i64 [[TMP7]], i64 [[TMP9]], i64 [[TMP1]], i64 [[TMP3]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP49:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP49]])
+// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP57]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP50:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP50]])
-// CHECK9-NEXT:    [[TMP51:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP51]]
+// CHECK9-NEXT:    [[TMP58:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP58]])
+// CHECK9-NEXT:    [[TMP59:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP59]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l86
@@ -1187,9 +1238,26 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 20)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l72.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l72.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l72([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1495,20 +1563,37 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP45:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_3]], align 8
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP45]], 1
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK11-NEXT:    [[TMP46:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l86.region_id, i32 5, i8** [[TMP38]], i8** [[TMP39]], i64* [[TMP40]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP47:%.*]] = icmp ne i32 [[TMP46]], 0
-// CHECK11-NEXT:    br i1 [[TMP47]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP38]], i8*** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP40]], i64** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l86.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP55:%.*]] = icmp ne i32 [[TMP54]], 0
+// CHECK11-NEXT:    br i1 [[TMP55]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l86(i32 [[TMP5]], i32 [[TMP7]], i32 [[TMP0]], i32 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP48:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP48]])
+// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP56]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP49]])
-// CHECK11-NEXT:    [[TMP50:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP50]]
+// CHECK11-NEXT:    [[TMP57:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP57]])
+// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP58]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l86
@@ -1859,9 +1944,26 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 20)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l72.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l72.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l72([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_simd_dist_schedule_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_simd_dist_schedule_codegen.cpp
index 20772a3..d02ddeb 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_simd_dist_schedule_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_simd_dist_schedule_codegen.cpp
@@ -166,10 +166,10 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -184,55 +184,106 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK1:       omp_offload.failed16:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK1:       omp_offload.cont15:
-// CHECK1-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP27]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK1:       omp_offload.cont17:
+// CHECK1-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i64 0, i64 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -777,10 +828,10 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -795,55 +846,106 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK3:       omp_offload.failed8:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK3:       omp_offload.cont8:
-// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK3:       omp_offload.cont9:
+// CHECK3-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK3-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK3:       omp_offload.failed14:
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK3:       omp_offload.cont15:
-// CHECK3-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP27]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -1627,14 +1729,14 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[M_CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES24:%.*]] = alloca [4 x i64], align 8
-// CHECK9-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
+// CHECK9-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -1693,140 +1795,191 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK9-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK9-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK9-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
-// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK9-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK9-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK9-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK9-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK9-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK9-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK9-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP58]])
-// CHECK9-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l113.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK9-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK9:       omp_offload.failed16:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l113(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK9:       omp_offload.cont17:
-// CHECK9-NEXT:    [[TMP61:%.*]] = load i32, i32* [[M]], align 4
-// CHECK9-NEXT:    [[CONV18:%.*]] = bitcast i64* [[M_CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP61]], i32* [[CONV18]], align 4
-// CHECK9-NEXT:    [[TMP62:%.*]] = load i64, i64* [[M_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP63]], i32* [[CONV20]], align 4
-// CHECK9-NEXT:    [[TMP64:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK9-NEXT:    [[TMP65:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP66:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES24]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP66]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
-// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP62]], i64* [[TMP68]], align 8
-// CHECK9-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP62]], i64* [[TMP70]], align 8
-// CHECK9-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP71]], align 8
-// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP64]], i64* [[TMP73]], align 8
-// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP64]], i64* [[TMP75]], align 8
-// CHECK9-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP76]], align 8
-// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP78]], align 8
-// CHECK9-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP80]], align 8
-// CHECK9-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP81]], align 8
-// CHECK9-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 8
-// CHECK9-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
-// CHECK9-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP85]], align 8
-// CHECK9-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 3
-// CHECK9-NEXT:    store i64 [[TMP65]], i64* [[TMP86]], align 8
-// CHECK9-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK9-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK9-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP66]])
+// CHECK9-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK9-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK9-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK9-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK9-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l113.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK9-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK9-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK9:       omp_offload.failed17:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l113(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK9:       omp_offload.cont18:
+// CHECK9-NEXT:    [[TMP77:%.*]] = load i32, i32* [[M]], align 4
+// CHECK9-NEXT:    [[CONV19:%.*]] = bitcast i64* [[M_CASTED]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP77]], i32* [[CONV19]], align 4
+// CHECK9-NEXT:    [[TMP78:%.*]] = load i64, i64* [[M_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP79]], i32* [[CONV21]], align 4
+// CHECK9-NEXT:    [[TMP80:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK9-NEXT:    [[TMP81:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP82:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP82]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
+// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP78]], i64* [[TMP84]], align 8
+// CHECK9-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP78]], i64* [[TMP86]], align 8
+// CHECK9-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
 // CHECK9-NEXT:    store i8* null, i8** [[TMP87]], align 8
-// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP91:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP91]], i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK9-NEXT:    [[TMP92:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK9-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[TMP92]], 0
-// CHECK9-NEXT:    [[DIV29:%.*]] = sdiv i32 [[SUB28]], 1
-// CHECK9-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[DIV29]], 1
-// CHECK9-NEXT:    store i32 [[SUB30]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK9-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK9-NEXT:    [[ADD31:%.*]] = add nsw i32 [[TMP93]], 1
-// CHECK9-NEXT:    [[TMP94:%.*]] = zext i32 [[ADD31]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP94]])
-// CHECK9-NEXT:    [[TMP95:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l118.region_id, i32 4, i8** [[TMP88]], i8** [[TMP89]], i64* [[TMP90]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP96:%.*]] = icmp ne i32 [[TMP95]], 0
-// CHECK9-NEXT:    br i1 [[TMP96]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
-// CHECK9:       omp_offload.failed32:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l118(i64 [[TMP62]], i64 [[TMP64]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
-// CHECK9:       omp_offload.cont33:
-// CHECK9-NEXT:    [[TMP97:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP97]])
+// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP80]], i64* [[TMP89]], align 8
+// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP80]], i64* [[TMP91]], align 8
+// CHECK9-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP92]], align 8
+// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP94]], align 8
+// CHECK9-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP96]], align 8
+// CHECK9-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP97]], align 8
+// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 8
+// CHECK9-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK9-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP101]], align 8
+// CHECK9-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 3
+// CHECK9-NEXT:    store i64 [[TMP81]], i64* [[TMP102]], align 8
+// CHECK9-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK9-NEXT:    store i8* null, i8** [[TMP103]], align 8
+// CHECK9-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP107:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP107]], i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK9-NEXT:    [[TMP108:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK9-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP108]], 0
+// CHECK9-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
+// CHECK9-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
+// CHECK9-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK9-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK9-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP109]], 1
+// CHECK9-NEXT:    [[TMP110:%.*]] = zext i32 [[ADD32]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP110]])
+// CHECK9-NEXT:    [[KERNEL_ARGS33:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP111]], align 4
+// CHECK9-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP112]], align 4
+// CHECK9-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP104]], i8*** [[TMP113]], align 8
+// CHECK9-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK9-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP106]], i64** [[TMP115]], align 8
+// CHECK9-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP116]], align 8
+// CHECK9-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP117]], align 8
+// CHECK9-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK9-NEXT:    [[TMP119:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l118.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]])
+// CHECK9-NEXT:    [[TMP120:%.*]] = icmp ne i32 [[TMP119]], 0
+// CHECK9-NEXT:    br i1 [[TMP120]], label [[OMP_OFFLOAD_FAILED34:%.*]], label [[OMP_OFFLOAD_CONT35:%.*]]
+// CHECK9:       omp_offload.failed34:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l118(i64 [[TMP78]], i64 [[TMP80]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT35]]
+// CHECK9:       omp_offload.cont35:
+// CHECK9-NEXT:    [[TMP121:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP121]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP98:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP98]])
-// CHECK9-NEXT:    [[TMP99:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP99]]
+// CHECK9-NEXT:    [[TMP122:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP122]])
+// CHECK9-NEXT:    [[TMP123:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP123]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108
@@ -2599,10 +2752,10 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[M_CASTED:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -2616,61 +2769,112 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l86.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l86.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l86([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l91.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK9-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK9:       omp_offload.failed5:
+// CHECK9-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l91.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK9-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK9-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK9:       omp_offload.failed6:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l91([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK9:       omp_offload.cont6:
-// CHECK9-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK9:       omp_offload.cont7:
+// CHECK9-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[M_CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP18]], i32* [[CONV]], align 4
-// CHECK9-NEXT:    [[TMP19:%.*]] = load i64, i64* [[M_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP19]], i64* [[TMP21]], align 8
-// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP19]], i64* [[TMP23]], align 8
-// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP24]], align 8
-// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP26]], align 8
-// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP28]], align 8
-// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP29]], align 8
-// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 [[TMP34]], i32* [[CONV]], align 4
+// CHECK9-NEXT:    [[TMP35:%.*]] = load i64, i64* [[M_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP35]], i64* [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP35]], i64* [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP44]], align 8
+// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l96.region_id, i32 2, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK9-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK9:       omp_offload.failed11:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l96(i64 [[TMP19]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK9:       omp_offload.cont12:
+// CHECK9-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 2, i32* [[TMP49]], align 4
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP46]], i8*** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 8
+// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l96.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK9-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK9-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK9:       omp_offload.failed13:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l96(i64 [[TMP35]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK9:       omp_offload.cont14:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -3243,14 +3447,14 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[M_CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED17:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS18:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES21:%.*]] = alloca [4 x i64], align 4
-// CHECK11-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_23:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
+// CHECK11-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -3308,139 +3512,190 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK11-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK11-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK11-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK11-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK11-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
-// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK11-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK11-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK11-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK11-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK11-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK11-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP59]])
-// CHECK11-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l113.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK11-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK11:       omp_offload.failed15:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l113(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK11:       omp_offload.cont16:
-// CHECK11-NEXT:    [[TMP62:%.*]] = load i32, i32* [[M]], align 4
-// CHECK11-NEXT:    store i32 [[TMP62]], i32* [[M_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP63:%.*]] = load i32, i32* [[M_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[N_CASTED17]], align 4
-// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[N_CASTED17]], align 4
-// CHECK11-NEXT:    [[TMP66:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP67:%.*]] = sext i32 [[TMP66]] to i64
-// CHECK11-NEXT:    [[TMP68:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES21]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP68]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
-// CHECK11-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP63]], i32* [[TMP70]], align 4
-// CHECK11-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP63]], i32* [[TMP72]], align 4
-// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP73]], align 4
-// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP65]], i32* [[TMP75]], align 4
-// CHECK11-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP65]], i32* [[TMP77]], align 4
-// CHECK11-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP78]], align 4
-// CHECK11-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP80]], align 4
-// CHECK11-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP82]], align 4
-// CHECK11-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP83]], align 4
-// CHECK11-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP85]], align 4
-// CHECK11-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 3
-// CHECK11-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP87]], align 4
-// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 3
-// CHECK11-NEXT:    store i64 [[TMP67]], i64* [[TMP88]], align 4
-// CHECK11-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK11-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP67]])
+// CHECK11-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK11-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK11-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK11-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK11-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l113.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK11-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK11-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK11:       omp_offload.failed16:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l113(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK11:       omp_offload.cont17:
+// CHECK11-NEXT:    [[TMP78:%.*]] = load i32, i32* [[M]], align 4
+// CHECK11-NEXT:    store i32 [[TMP78]], i32* [[M_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP79:%.*]] = load i32, i32* [[M_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP80]], i32* [[N_CASTED18]], align 4
+// CHECK11-NEXT:    [[TMP81:%.*]] = load i32, i32* [[N_CASTED18]], align 4
+// CHECK11-NEXT:    [[TMP82:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP83:%.*]] = sext i32 [[TMP82]] to i64
+// CHECK11-NEXT:    [[TMP84:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP84]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
+// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP79]], i32* [[TMP86]], align 4
+// CHECK11-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP88:%.*]] = bitcast i8** [[TMP87]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP79]], i32* [[TMP88]], align 4
+// CHECK11-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
 // CHECK11-NEXT:    store i8* null, i8** [[TMP89]], align 4
-// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP93:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP93]], i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK11-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK11-NEXT:    [[SUB25:%.*]] = sub nsw i32 [[TMP94]], 0
-// CHECK11-NEXT:    [[DIV26:%.*]] = sdiv i32 [[SUB25]], 1
-// CHECK11-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[DIV26]], 1
-// CHECK11-NEXT:    store i32 [[SUB27]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[ADD28:%.*]] = add nsw i32 [[TMP95]], 1
-// CHECK11-NEXT:    [[TMP96:%.*]] = zext i32 [[ADD28]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP96]])
-// CHECK11-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l118.region_id, i32 4, i8** [[TMP90]], i8** [[TMP91]], i64* [[TMP92]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
-// CHECK11-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
-// CHECK11:       omp_offload.failed29:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l118(i32 [[TMP63]], i32 [[TMP65]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
-// CHECK11:       omp_offload.cont30:
-// CHECK11-NEXT:    [[TMP99:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP99]])
+// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP81]], i32* [[TMP91]], align 4
+// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP81]], i32* [[TMP93]], align 4
+// CHECK11-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP94]], align 4
+// CHECK11-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP96]], align 4
+// CHECK11-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP98]], align 4
+// CHECK11-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP99]], align 4
+// CHECK11-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP101]], align 4
+// CHECK11-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK11-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP103]], align 4
+// CHECK11-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 3
+// CHECK11-NEXT:    store i64 [[TMP83]], i64* [[TMP104]], align 4
+// CHECK11-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK11-NEXT:    store i8* null, i8** [[TMP105]], align 4
+// CHECK11-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP109:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP109]], i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK11-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK11-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP110]], 0
+// CHECK11-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
+// CHECK11-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
+// CHECK11-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP111]], 1
+// CHECK11-NEXT:    [[TMP112:%.*]] = zext i32 [[ADD29]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP112]])
+// CHECK11-NEXT:    [[KERNEL_ARGS30:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP113]], align 4
+// CHECK11-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP114]], align 4
+// CHECK11-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 4
+// CHECK11-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK11-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP108]], i64** [[TMP117]], align 4
+// CHECK11-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP118]], align 4
+// CHECK11-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP119]], align 4
+// CHECK11-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK11-NEXT:    [[TMP121:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l118.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]])
+// CHECK11-NEXT:    [[TMP122:%.*]] = icmp ne i32 [[TMP121]], 0
+// CHECK11-NEXT:    br i1 [[TMP122]], label [[OMP_OFFLOAD_FAILED31:%.*]], label [[OMP_OFFLOAD_CONT32:%.*]]
+// CHECK11:       omp_offload.failed31:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l118(i32 [[TMP79]], i32 [[TMP81]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT32]]
+// CHECK11:       omp_offload.cont32:
+// CHECK11-NEXT:    [[TMP123:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP123]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP100:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP100]])
-// CHECK11-NEXT:    [[TMP101:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP101]]
+// CHECK11-NEXT:    [[TMP124:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP124]])
+// CHECK11-NEXT:    [[TMP125:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP125]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l108
@@ -4190,10 +4445,10 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 4
 // CHECK11-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[M_CASTED:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -4207,60 +4462,111 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l86.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l86.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l86([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK11-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l91.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK11-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK11:       omp_offload.failed5:
+// CHECK11-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l91.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK11-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK11-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK11:       omp_offload.failed6:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l91([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK11:       omp_offload.cont6:
-// CHECK11-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK11-NEXT:    store i32 [[TMP18]], i32* [[M_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP19:%.*]] = load i32, i32* [[M_CASTED]], align 4
-// CHECK11-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP19]], i32* [[TMP21]], align 4
-// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP19]], i32* [[TMP23]], align 4
-// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP24]], align 4
-// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP26]], align 4
-// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP28]], align 4
-// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP29]], align 4
-// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK11:       omp_offload.cont7:
+// CHECK11-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK11-NEXT:    store i32 [[TMP34]], i32* [[M_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = load i32, i32* [[M_CASTED]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP35]], i32* [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP35]], i32* [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP44]], align 4
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP45]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l96.region_id, i32 2, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK11-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK11:       omp_offload.failed11:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l96(i32 [[TMP19]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK11:       omp_offload.cont12:
+// CHECK11-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 2, i32* [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP46]], i8*** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP55]], align 4
+// CHECK11-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l96.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK11-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK11-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK11:       omp_offload.failed13:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l96(i32 [[TMP35]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK11:       omp_offload.cont14:
 // CHECK11-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_simd_firstprivate_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_simd_firstprivate_codegen.cpp
index e3fed04..e4be256 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_simd_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_simd_firstprivate_codegen.cpp
@@ -354,9 +354,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l101.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK1-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l101.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK1-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l101(i64 [[TMP1]], [2 x i32]* @vec, [2 x %struct.S]* @s_arr, %struct.S* @var, i64 [[TMP3]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -781,27 +798,44 @@ int main() {
 // CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK1-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK1-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP32]]
+// CHECK1-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN2StC2Ev
@@ -1444,9 +1478,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l101.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK3-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l101.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK3-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l101(i32 [[TMP1]], [2 x i32]* @vec, [2 x %struct.S]* @s_arr, %struct.S* @var, i32 [[TMP3]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1854,27 +1905,44 @@ int main() {
 // CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK3-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.6, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.7, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK3-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP32]]
+// CHECK3-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN2StC2Ev

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_simd_if_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_simd_if_codegen.cpp
index 569124e..4e7149f 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_simd_if_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_simd_if_codegen.cpp
@@ -116,21 +116,55 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43() #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret void
 //
 //
@@ -453,55 +487,106 @@ int main() {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i8, align 1
-// CHECK1-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l84.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l84() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
+// CHECK1-NEXT:    [[TMP20:%.*]] = load i32, i32* @Arg, align 4
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP13:%.*]] = load i32, i32* @Arg, align 4
-// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP13]], 0
+// CHECK1-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP29:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP29]], 0
 // CHECK1-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP14:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP14]] to i1
-// CHECK1-NEXT:    [[TMP15:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP30:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL5:%.*]] = trunc i8 [[TMP30]] to i1
+// CHECK1-NEXT:    [[TMP31:%.*]] = select i1 [[TOBOOL5]], i32 0, i32 1
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP15]])
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK1:       omp_offload.failed6:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK1:       omp_offload.cont7:
-// CHECK1-NEXT:    [[TMP18:%.*]] = load i32, i32* @Arg, align 4
-// CHECK1-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP18]])
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP31]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK1-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[TMP42:%.*]] = load i32, i32* @Arg, align 4
+// CHECK1-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP42]])
 // CHECK1-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -1009,53 +1094,104 @@ int main() {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i8, align 1
-// CHECK1-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l57() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
+// CHECK1-NEXT:    [[TMP20:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP13:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
-// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP13]], 0
+// CHECK1-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK1-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP29:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK1-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP29]], 0
 // CHECK1-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK1-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP14:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP14]] to i1
-// CHECK1-NEXT:    [[TMP15:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP30:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TOBOOL5:%.*]] = trunc i8 [[TMP30]] to i1
+// CHECK1-NEXT:    [[TMP31:%.*]] = select i1 [[TOBOOL5]], i32 0, i32 1
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP15]])
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK1:       omp_offload.failed6:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK1:       omp_offload.cont7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP39]], align 8
+// CHECK1-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP31]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK1-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
 // CHECK1-NEXT:    ret i32 0
 //
 //
@@ -1565,21 +1701,55 @@ int main() {
 // CHECK3-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43() #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK3-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK3:       omp_offload.failed2:
+// CHECK3-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK3-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK3-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK3-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK3:       omp_offload.failed3:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47() #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK3:       omp_offload.cont3:
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK3:       omp_offload.cont4:
 // CHECK3-NEXT:    ret void
 //
 //
@@ -1902,55 +2072,106 @@ int main() {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i8, align 1
-// CHECK3-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l84.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK3-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK3:       omp_offload.failed2:
+// CHECK3-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK3-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK3-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK3-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK3:       omp_offload.failed3:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l84() #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK3:       omp_offload.cont3:
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK3:       omp_offload.cont4:
+// CHECK3-NEXT:    [[TMP20:%.*]] = load i32, i32* @Arg, align 4
 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK3-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK3-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK3-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP13:%.*]] = load i32, i32* @Arg, align 4
-// CHECK3-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP13]], 0
+// CHECK3-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK3-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK3-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP29:%.*]] = load i32, i32* @Arg, align 4
+// CHECK3-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP29]], 0
 // CHECK3-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK3-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK3-NEXT:    [[TMP14:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK3-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP14]] to i1
-// CHECK3-NEXT:    [[TMP15:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP30:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK3-NEXT:    [[TOBOOL5:%.*]] = trunc i8 [[TMP30]] to i1
+// CHECK3-NEXT:    [[TMP31:%.*]] = select i1 [[TOBOOL5]], i32 0, i32 1
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP15]])
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK3:       omp_offload.failed6:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK3:       omp_offload.cont7:
-// CHECK3-NEXT:    [[TMP18:%.*]] = load i32, i32* @Arg, align 4
-// CHECK3-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP18]])
+// CHECK3-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP34]], align 8
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP35]], align 8
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP37]], align 8
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP39]], align 8
+// CHECK3-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP31]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK3-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK3-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK3:       omp_offload.failed8:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK3:       omp_offload.cont9:
+// CHECK3-NEXT:    [[TMP42:%.*]] = load i32, i32* @Arg, align 4
+// CHECK3-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP42]])
 // CHECK3-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -2692,53 +2913,104 @@ int main() {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
 // CHECK3-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i8, align 1
-// CHECK3-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l57() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK3-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK3:       omp_offload.failed2:
+// CHECK3-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK3-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK3-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK3-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK3:       omp_offload.failed3:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62() #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK3:       omp_offload.cont3:
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK3:       omp_offload.cont4:
+// CHECK3-NEXT:    [[TMP20:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
 // CHECK3-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK3-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK3-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK3-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP13:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
-// CHECK3-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP13]], 0
+// CHECK3-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK3-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK3-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP29:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK3-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP29]], 0
 // CHECK3-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK3-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK3-NEXT:    [[TMP14:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK3-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP14]] to i1
-// CHECK3-NEXT:    [[TMP15:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP30:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK3-NEXT:    [[TOBOOL5:%.*]] = trunc i8 [[TMP30]] to i1
+// CHECK3-NEXT:    [[TMP31:%.*]] = select i1 [[TOBOOL5]], i32 0, i32 1
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP15]])
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK3:       omp_offload.failed6:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK3:       omp_offload.cont7:
+// CHECK3-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP34]], align 8
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP35]], align 8
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP37]], align 8
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP39]], align 8
+// CHECK3-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP31]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK3-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK3-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK3:       omp_offload.failed8:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK3:       omp_offload.cont9:
 // CHECK3-NEXT:    ret i32 0
 //
 //
@@ -3839,21 +4111,55 @@ int main() {
 // CHECK9-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43() #[[ATTR2:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK9-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK9:       omp_offload.failed2:
+// CHECK9-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK9-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK9-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK9:       omp_offload.failed3:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47() #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK9:       omp_offload.cont3:
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK9:       omp_offload.cont4:
 // CHECK9-NEXT:    ret void
 //
 //
@@ -4176,55 +4482,106 @@ int main() {
 // CHECK9-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i8, align 1
-// CHECK9-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76() #[[ATTR2]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l84.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK9-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK9:       omp_offload.failed2:
+// CHECK9-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK9-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK9-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK9:       omp_offload.failed3:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l84() #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK9:       omp_offload.cont3:
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK9:       omp_offload.cont4:
+// CHECK9-NEXT:    [[TMP20:%.*]] = load i32, i32* @Arg, align 4
 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK9-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP13:%.*]] = load i32, i32* @Arg, align 4
-// CHECK9-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP13]], 0
+// CHECK9-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK9-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP29:%.*]] = load i32, i32* @Arg, align 4
+// CHECK9-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP29]], 0
 // CHECK9-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK9-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK9-NEXT:    [[TMP14:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK9-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP14]] to i1
-// CHECK9-NEXT:    [[TMP15:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP30:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK9-NEXT:    [[TOBOOL5:%.*]] = trunc i8 [[TMP30]] to i1
+// CHECK9-NEXT:    [[TMP31:%.*]] = select i1 [[TOBOOL5]], i32 0, i32 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP15]])
-// CHECK9-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK9-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK9:       omp_offload.failed6:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK9:       omp_offload.cont7:
-// CHECK9-NEXT:    [[TMP18:%.*]] = load i32, i32* @Arg, align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP18]])
+// CHECK9-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP31]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK9-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK9-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK9:       omp_offload.failed8:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK9:       omp_offload.cont9:
+// CHECK9-NEXT:    [[TMP42:%.*]] = load i32, i32* @Arg, align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP42]])
 // CHECK9-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -4732,53 +5089,104 @@ int main() {
 // CHECK9-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i8, align 1
-// CHECK9-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK9-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l57() #[[ATTR2]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK9-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK9:       omp_offload.failed2:
+// CHECK9-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK9-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK9-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK9-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK9:       omp_offload.failed3:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62() #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK9:       omp_offload.cont3:
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK9:       omp_offload.cont4:
+// CHECK9-NEXT:    [[TMP20:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
 // CHECK9-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK9-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP13:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
-// CHECK9-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP13]], 0
+// CHECK9-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK9-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP29:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK9-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP29]], 0
 // CHECK9-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK9-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK9-NEXT:    [[TMP14:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK9-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP14]] to i1
-// CHECK9-NEXT:    [[TMP15:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP30:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK9-NEXT:    [[TOBOOL5:%.*]] = trunc i8 [[TMP30]] to i1
+// CHECK9-NEXT:    [[TMP31:%.*]] = select i1 [[TOBOOL5]], i32 0, i32 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK9-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP15]])
-// CHECK9-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK9-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK9:       omp_offload.failed6:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK9:       omp_offload.cont7:
+// CHECK9-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP31]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK9-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK9-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK9:       omp_offload.failed8:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK9:       omp_offload.cont9:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -5288,21 +5696,55 @@ int main() {
 // CHECK11-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK11-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK11-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK11-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l43() #[[ATTR2:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK11-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK11:       omp_offload.failed2:
+// CHECK11-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK11-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK11-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK11-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK11-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK11:       omp_offload.failed3:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9gtid_testv_l47() #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK11:       omp_offload.cont3:
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK11:       omp_offload.cont4:
 // CHECK11-NEXT:    ret void
 //
 //
@@ -5625,55 +6067,106 @@ int main() {
 // CHECK11-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i8, align 1
-// CHECK11-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK11-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK11-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK11-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l76() #[[ATTR2]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l84.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK11-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK11:       omp_offload.failed2:
+// CHECK11-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK11-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK11-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK11-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK11-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK11:       omp_offload.failed3:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l84() #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK11:       omp_offload.cont3:
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* @Arg, align 4
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK11:       omp_offload.cont4:
+// CHECK11-NEXT:    [[TMP20:%.*]] = load i32, i32* @Arg, align 4
 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK11-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK11-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK11-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP13:%.*]] = load i32, i32* @Arg, align 4
-// CHECK11-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP13]], 0
+// CHECK11-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK11-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK11-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP29:%.*]] = load i32, i32* @Arg, align 4
+// CHECK11-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP29]], 0
 // CHECK11-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK11-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK11-NEXT:    [[TMP14:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK11-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP14]] to i1
-// CHECK11-NEXT:    [[TMP15:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP30:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK11-NEXT:    [[TOBOOL5:%.*]] = trunc i8 [[TMP30]] to i1
+// CHECK11-NEXT:    [[TMP31:%.*]] = select i1 [[TOBOOL5]], i32 0, i32 1
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP15]])
-// CHECK11-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK11-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK11:       omp_offload.failed6:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK11:       omp_offload.cont7:
-// CHECK11-NEXT:    [[TMP18:%.*]] = load i32, i32* @Arg, align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP18]])
+// CHECK11-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP34]], align 8
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP35]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP37]], align 8
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP39]], align 8
+// CHECK11-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP31]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK11-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK11-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK11:       omp_offload.failed8:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l92(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK11:       omp_offload.cont9:
+// CHECK11-NEXT:    [[TMP42:%.*]] = load i32, i32* @Arg, align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiEiT_(i32 noundef [[TMP42]])
 // CHECK11-NEXT:    ret i32 [[CALL]]
 //
 //
@@ -6415,53 +6908,104 @@ int main() {
 // CHECK11-NEXT:    [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
 // CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i8, align 1
-// CHECK11-NEXT:    [[_TMP5:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 [[ARG]], i32* [[ARG_ADDR]], align 4
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l57.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK11-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK11-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l57.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK11-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l57() #[[ATTR2]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK11-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK11:       omp_offload.failed2:
+// CHECK11-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK11-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK11-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK11-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK11-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK11:       omp_offload.failed3:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l62() #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK11:       omp_offload.cont3:
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK11:       omp_offload.cont4:
+// CHECK11-NEXT:    [[TMP20:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
 // CHECK11-NEXT:    [[CONV:%.*]] = bitcast i64* [[ARG_CASTED]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP4]], i32* [[CONV]], align 4
-// CHECK11-NEXT:    [[TMP5:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
-// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i64*
-// CHECK11-NEXT:    store i64 [[TMP5]], i64* [[TMP7]], align 8
-// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP9:%.*]] = bitcast i8** [[TMP8]] to i64*
-// CHECK11-NEXT:    store i64 [[TMP5]], i64* [[TMP9]], align 8
-// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP10]], align 8
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP13:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
-// CHECK11-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP13]], 0
+// CHECK11-NEXT:    store i32 [[TMP20]], i32* [[CONV]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = load i64, i64* [[ARG_CASTED]], align 8
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
+// CHECK11-NEXT:    store i64 [[TMP21]], i64* [[TMP23]], align 8
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP25:%.*]] = bitcast i8** [[TMP24]] to i64*
+// CHECK11-NEXT:    store i64 [[TMP21]], i64* [[TMP25]], align 8
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP26]], align 8
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP29:%.*]] = load i32, i32* [[ARG_ADDR]], align 4
+// CHECK11-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP29]], 0
 // CHECK11-NEXT:    [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8
 // CHECK11-NEXT:    store i8 [[FROMBOOL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK11-NEXT:    [[TMP14:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK11-NEXT:    [[TOBOOL4:%.*]] = trunc i8 [[TMP14]] to i1
-// CHECK11-NEXT:    [[TMP15:%.*]] = select i1 [[TOBOOL4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP30:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK11-NEXT:    [[TOBOOL5:%.*]] = trunc i8 [[TMP30]] to i1
+// CHECK11-NEXT:    [[TMP31:%.*]] = select i1 [[TOBOOL5]], i32 0, i32 1
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK11-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67.region_id, i32 1, i8** [[TMP11]], i8** [[TMP12]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP15]])
-// CHECK11-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK11-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
-// CHECK11:       omp_offload.failed6:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67(i64 [[TMP5]]) #[[ATTR2]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
-// CHECK11:       omp_offload.cont7:
+// CHECK11-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP34]], align 8
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP35]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.17, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.18, i32 0, i32 0), i64** [[TMP37]], align 8
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP39]], align 8
+// CHECK11-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP31]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK11-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK11-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK11:       omp_offload.failed8:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiEiT__l67(i64 [[TMP21]]) #[[ATTR2]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK11:       omp_offload.cont9:
 // CHECK11-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_simd_lastprivate_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_simd_lastprivate_codegen.cpp
index ecb28ea..011c8d2 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_simd_lastprivate_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_simd_lastprivate_codegen.cpp
@@ -870,9 +870,26 @@ int main() {
 // CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK9-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK9-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -880,18 +897,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done3:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP39]]
+// CHECK9-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1345,27 +1362,44 @@ int main() {
 // CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK9-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK9-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP32]]
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1895,9 +1929,26 @@ int main() {
 // CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK11-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK11-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1905,18 +1956,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP39]]
+// CHECK11-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -2361,27 +2412,44 @@ int main() {
 // CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK11-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK11-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP32]]
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_simd_num_threads_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_simd_num_threads_codegen.cpp
index e6f2221..3ec9c79 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_simd_num_threads_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_simd_num_threads_codegen.cpp
@@ -97,68 +97,102 @@ int main() {
 // CHECK1:       invoke.cont:
 // CHECK1-NEXT:    store i8 [[CALL]], i8* [[A]], align 1
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l50.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 2)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 2, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l50.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l50() #[[ATTR6:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       lpad:
-// CHECK1-NEXT:    [[TMP2:%.*]] = landingpad { i8*, i32 }
+// CHECK1-NEXT:    [[TMP10:%.*]] = landingpad { i8*, i32 }
 // CHECK1-NEXT:    cleanup
-// CHECK1-NEXT:    [[TMP3:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 0
-// CHECK1-NEXT:    store i8* [[TMP3]], i8** [[EXN_SLOT]], align 8
-// CHECK1-NEXT:    [[TMP4:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 1
-// CHECK1-NEXT:    store i32 [[TMP4]], i32* [[EHSELECTOR_SLOT]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 0
+// CHECK1-NEXT:    store i8* [[TMP11]], i8** [[EXN_SLOT]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 1
+// CHECK1-NEXT:    store i32 [[TMP12]], i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK1-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
 // CHECK1-NEXT:    br label [[EH_RESUME:%.*]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP5:%.*]] = load i8, i8* [[A]], align 1
+// CHECK1-NEXT:    [[TMP13:%.*]] = load i8, i8* [[A]], align 1
 // CHECK1-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i8*
-// CHECK1-NEXT:    store i8 [[TMP5]], i8* [[CONV]], align 1
-// CHECK1-NEXT:    [[TMP6:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP8]], align 8
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP6]], i64* [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP11]], align 8
-// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP14:%.*]] = load i8, i8* [[A]], align 1
-// CHECK1-NEXT:    store i8 [[TMP14]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP15:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP16:%.*]] = zext i8 [[TMP15]] to i32
+// CHECK1-NEXT:    store i8 [[TMP13]], i8* [[CONV]], align 1
+// CHECK1-NEXT:    [[TMP14:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP14]], i64* [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP22:%.*]] = load i8, i8* [[A]], align 1
+// CHECK1-NEXT:    store i8 [[TMP22]], i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TMP23:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TMP24:%.*]] = zext i8 [[TMP23]] to i32
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55.region_id, i32 1, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP16]])
-// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
-// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55(i64 [[TMP6]]) #[[ATTR6]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
-// CHECK1-NEXT:    [[TMP19:%.*]] = load i8, i8* [[A]], align 1
-// CHECK1-NEXT:    [[CONV4:%.*]] = sext i8 [[TMP19]] to i32
-// CHECK1-NEXT:    [[CALL6:%.*]] = invoke noundef signext i32 @_Z5tmainIcLi5EEiv()
-// CHECK1-NEXT:    to label [[INVOKE_CONT5:%.*]] unwind label [[LPAD]]
-// CHECK1:       invoke.cont5:
-// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV4]], [[CALL6]]
-// CHECK1-NEXT:    [[CALL8:%.*]] = invoke noundef signext i32 @_Z5tmainI1SLi1EEiv()
-// CHECK1-NEXT:    to label [[INVOKE_CONT7:%.*]] unwind label [[LPAD]]
-// CHECK1:       invoke.cont7:
-// CHECK1-NEXT:    [[ADD9:%.*]] = add nsw i32 [[ADD]], [[CALL8]]
-// CHECK1-NEXT:    store i32 [[ADD9]], i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP20]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP21]], i8*** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP24]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
+// CHECK1-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55(i64 [[TMP14]]) #[[ATTR6]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
+// CHECK1-NEXT:    [[TMP35:%.*]] = load i8, i8* [[A]], align 1
+// CHECK1-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP35]] to i32
+// CHECK1-NEXT:    [[CALL7:%.*]] = invoke noundef signext i32 @_Z5tmainIcLi5EEiv()
+// CHECK1-NEXT:    to label [[INVOKE_CONT6:%.*]] unwind label [[LPAD]]
+// CHECK1:       invoke.cont6:
+// CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV5]], [[CALL7]]
+// CHECK1-NEXT:    [[CALL9:%.*]] = invoke noundef signext i32 @_Z5tmainI1SLi1EEiv()
+// CHECK1-NEXT:    to label [[INVOKE_CONT8:%.*]] unwind label [[LPAD]]
+// CHECK1:       invoke.cont8:
+// CHECK1-NEXT:    [[ADD10:%.*]] = add nsw i32 [[ADD]], [[CALL9]]
+// CHECK1-NEXT:    store i32 [[ADD10]], i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
-// CHECK1-NEXT:    [[TMP20:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP20]]
+// CHECK1-NEXT:    [[TMP36:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP36]]
 // CHECK1:       eh.resume:
 // CHECK1-NEXT:    [[EXN:%.*]] = load i8*, i8** [[EXN_SLOT]], align 8
 // CHECK1-NEXT:    [[SEL:%.*]] = load i32, i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK1-NEXT:    [[LPAD_VAL:%.*]] = insertvalue { i8*, i32 } undef, i8* [[EXN]], 0
-// CHECK1-NEXT:    [[LPAD_VAL10:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
-// CHECK1-NEXT:    resume { i8*, i32 } [[LPAD_VAL10]]
+// CHECK1-NEXT:    [[LPAD_VAL11:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
+// CHECK1-NEXT:    resume { i8*, i32 } [[LPAD_VAL11]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN1SC1El
@@ -537,21 +571,55 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l36.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 5)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 5, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l36() #[[ATTR6]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l40.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 23)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 23, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l40() #[[ATTR6]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret i32 0
 //
 //
@@ -563,9 +631,26 @@ int main() {
 // CHECK1-NEXT:    [[REF_TMP:%.*]] = alloca [[STRUCT_S:%.*]], align 8
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l36.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l36() #[[ATTR6]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -576,22 +661,39 @@ int main() {
 // CHECK1-NEXT:    [[CALL:%.*]] = call noundef signext i8 @_ZN1ScvcEv(%struct.S* noundef nonnull align 8 dereferenceable(24) [[REF_TMP]])
 // CHECK1-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[REF_TMP]]) #[[ATTR6]]
 // CHECK1-NEXT:    store i8 [[CALL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP2:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK1-NEXT:    [[TMP3:%.*]] = zext i8 [[TMP2]] to i32
+// CHECK1-NEXT:    [[TMP10:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK1-NEXT:    [[TMP11:%.*]] = zext i8 [[TMP10]] to i32
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK1-NEXT:    [[TMP4:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l40.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 [[TMP3]])
-// CHECK1-NEXT:    [[TMP5:%.*]] = icmp ne i32 [[TMP4]], 0
-// CHECK1-NEXT:    br i1 [[TMP5]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP13]], align 4
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP19]], align 8
+// CHECK1-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP11]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
+// CHECK1-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l40() #[[ATTR6]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    ret i32 0
 // CHECK1:       terminate.lpad:
-// CHECK1-NEXT:    [[TMP6:%.*]] = landingpad { i8*, i32 }
+// CHECK1-NEXT:    [[TMP22:%.*]] = landingpad { i8*, i32 }
 // CHECK1-NEXT:    catch i8* null
-// CHECK1-NEXT:    [[TMP7:%.*]] = extractvalue { i8*, i32 } [[TMP6]], 0
-// CHECK1-NEXT:    call void @__clang_call_terminate(i8* [[TMP7]]) #[[ATTR9]]
+// CHECK1-NEXT:    [[TMP23:%.*]] = extractvalue { i8*, i32 } [[TMP22]], 0
+// CHECK1-NEXT:    call void @__clang_call_terminate(i8* [[TMP23]]) #[[ATTR9]]
 // CHECK1-NEXT:    unreachable
 //
 //
@@ -1674,68 +1776,102 @@ int main() {
 // CHECK5:       invoke.cont:
 // CHECK5-NEXT:    store i8 [[CALL]], i8* [[A]], align 1
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l50.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 2)
-// CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK5-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK5-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK5-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK5-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK5-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 2, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l50.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK5-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l50() #[[ATTR6:[0-9]+]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       lpad:
-// CHECK5-NEXT:    [[TMP2:%.*]] = landingpad { i8*, i32 }
+// CHECK5-NEXT:    [[TMP10:%.*]] = landingpad { i8*, i32 }
 // CHECK5-NEXT:    cleanup
-// CHECK5-NEXT:    [[TMP3:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 0
-// CHECK5-NEXT:    store i8* [[TMP3]], i8** [[EXN_SLOT]], align 8
-// CHECK5-NEXT:    [[TMP4:%.*]] = extractvalue { i8*, i32 } [[TMP2]], 1
-// CHECK5-NEXT:    store i32 [[TMP4]], i32* [[EHSELECTOR_SLOT]], align 4
+// CHECK5-NEXT:    [[TMP11:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 0
+// CHECK5-NEXT:    store i8* [[TMP11]], i8** [[EXN_SLOT]], align 8
+// CHECK5-NEXT:    [[TMP12:%.*]] = extractvalue { i8*, i32 } [[TMP10]], 1
+// CHECK5-NEXT:    store i32 [[TMP12]], i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK5-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
 // CHECK5-NEXT:    br label [[EH_RESUME:%.*]]
 // CHECK5:       omp_offload.cont:
-// CHECK5-NEXT:    [[TMP5:%.*]] = load i8, i8* [[A]], align 1
+// CHECK5-NEXT:    [[TMP13:%.*]] = load i8, i8* [[A]], align 1
 // CHECK5-NEXT:    [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i8*
-// CHECK5-NEXT:    store i8 [[TMP5]], i8* [[CONV]], align 1
-// CHECK5-NEXT:    [[TMP6:%.*]] = load i64, i64* [[A_CASTED]], align 8
-// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP6]], i64* [[TMP8]], align 8
-// CHECK5-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to i64*
-// CHECK5-NEXT:    store i64 [[TMP6]], i64* [[TMP10]], align 8
-// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP11]], align 8
-// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP14:%.*]] = load i8, i8* [[A]], align 1
-// CHECK5-NEXT:    store i8 [[TMP14]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK5-NEXT:    [[TMP15:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK5-NEXT:    [[TMP16:%.*]] = zext i8 [[TMP15]] to i32
+// CHECK5-NEXT:    store i8 [[TMP13]], i8* [[CONV]], align 1
+// CHECK5-NEXT:    [[TMP14:%.*]] = load i64, i64* [[A_CASTED]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP16:%.*]] = bitcast i8** [[TMP15]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP14]], i64* [[TMP16]], align 8
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to i64*
+// CHECK5-NEXT:    store i64 [[TMP14]], i64* [[TMP18]], align 8
+// CHECK5-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP19]], align 8
+// CHECK5-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP22:%.*]] = load i8, i8* [[A]], align 1
+// CHECK5-NEXT:    store i8 [[TMP22]], i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK5-NEXT:    [[TMP23:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK5-NEXT:    [[TMP24:%.*]] = zext i8 [[TMP23]] to i32
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55.region_id, i32 1, i8** [[TMP12]], i8** [[TMP13]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 [[TMP16]])
-// CHECK5-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
-// CHECK5-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK5:       omp_offload.failed2:
-// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55(i64 [[TMP6]]) #[[ATTR6]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK5:       omp_offload.cont3:
-// CHECK5-NEXT:    [[TMP19:%.*]] = load i8, i8* [[A]], align 1
-// CHECK5-NEXT:    [[CONV4:%.*]] = sext i8 [[TMP19]] to i32
-// CHECK5-NEXT:    [[CALL6:%.*]] = invoke noundef signext i32 @_Z5tmainIcLi5EEiv()
-// CHECK5-NEXT:    to label [[INVOKE_CONT5:%.*]] unwind label [[LPAD]]
-// CHECK5:       invoke.cont5:
-// CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV4]], [[CALL6]]
-// CHECK5-NEXT:    [[CALL8:%.*]] = invoke noundef signext i32 @_Z5tmainI1SLi1EEiv()
-// CHECK5-NEXT:    to label [[INVOKE_CONT7:%.*]] unwind label [[LPAD]]
-// CHECK5:       invoke.cont7:
-// CHECK5-NEXT:    [[ADD9:%.*]] = add nsw i32 [[ADD]], [[CALL8]]
-// CHECK5-NEXT:    store i32 [[ADD9]], i32* [[RETVAL]], align 4
+// CHECK5-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP26]], align 4
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP20]], i8*** [[TMP27]], align 8
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP21]], i8*** [[TMP28]], align 8
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK5-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP30]], align 8
+// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP32]], align 8
+// CHECK5-NEXT:    [[TMP33:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP24]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK5-NEXT:    [[TMP34:%.*]] = icmp ne i32 [[TMP33]], 0
+// CHECK5-NEXT:    br i1 [[TMP34]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK5:       omp_offload.failed3:
+// CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l55(i64 [[TMP14]]) #[[ATTR6]]
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK5:       omp_offload.cont4:
+// CHECK5-NEXT:    [[TMP35:%.*]] = load i8, i8* [[A]], align 1
+// CHECK5-NEXT:    [[CONV5:%.*]] = sext i8 [[TMP35]] to i32
+// CHECK5-NEXT:    [[CALL7:%.*]] = invoke noundef signext i32 @_Z5tmainIcLi5EEiv()
+// CHECK5-NEXT:    to label [[INVOKE_CONT6:%.*]] unwind label [[LPAD]]
+// CHECK5:       invoke.cont6:
+// CHECK5-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV5]], [[CALL7]]
+// CHECK5-NEXT:    [[CALL9:%.*]] = invoke noundef signext i32 @_Z5tmainI1SLi1EEiv()
+// CHECK5-NEXT:    to label [[INVOKE_CONT8:%.*]] unwind label [[LPAD]]
+// CHECK5:       invoke.cont8:
+// CHECK5-NEXT:    [[ADD10:%.*]] = add nsw i32 [[ADD]], [[CALL9]]
+// CHECK5-NEXT:    store i32 [[ADD10]], i32* [[RETVAL]], align 4
 // CHECK5-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR6]]
-// CHECK5-NEXT:    [[TMP20:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP20]]
+// CHECK5-NEXT:    [[TMP36:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP36]]
 // CHECK5:       eh.resume:
 // CHECK5-NEXT:    [[EXN:%.*]] = load i8*, i8** [[EXN_SLOT]], align 8
 // CHECK5-NEXT:    [[SEL:%.*]] = load i32, i32* [[EHSELECTOR_SLOT]], align 4
 // CHECK5-NEXT:    [[LPAD_VAL:%.*]] = insertvalue { i8*, i32 } undef, i8* [[EXN]], 0
-// CHECK5-NEXT:    [[LPAD_VAL10:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
-// CHECK5-NEXT:    resume { i8*, i32 } [[LPAD_VAL10]]
+// CHECK5-NEXT:    [[LPAD_VAL11:%.*]] = insertvalue { i8*, i32 } [[LPAD_VAL]], i32 [[SEL]], 1
+// CHECK5-NEXT:    resume { i8*, i32 } [[LPAD_VAL11]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@_ZN1SC1El
@@ -2114,21 +2250,55 @@ int main() {
 // CHECK5-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l36.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 5)
-// CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK5-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK5-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK5-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK5-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK5-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 5, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK5-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l36() #[[ATTR6]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l40.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 23)
-// CHECK5-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK5-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK5:       omp_offload.failed2:
+// CHECK5-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK5-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK5-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 23, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK5-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK5-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK5:       omp_offload.failed3:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIcLi5EEiv_l40() #[[ATTR6]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK5:       omp_offload.cont3:
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK5:       omp_offload.cont4:
 // CHECK5-NEXT:    ret i32 0
 //
 //
@@ -2140,9 +2310,26 @@ int main() {
 // CHECK5-NEXT:    [[REF_TMP:%.*]] = alloca [[STRUCT_S:%.*]], align 8
 // CHECK5-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l36.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK5-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK5-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK5-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK5-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK5-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK5-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK5-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK5-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l36() #[[ATTR6]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -2153,22 +2340,39 @@ int main() {
 // CHECK5-NEXT:    [[CALL:%.*]] = call noundef signext i8 @_ZN1ScvcEv(%struct.S* noundef nonnull align 8 dereferenceable(24) [[REF_TMP]])
 // CHECK5-NEXT:    call void @_ZN1SD1Ev(%struct.S* noundef nonnull align 8 dereferenceable(24) [[REF_TMP]]) #[[ATTR6]]
 // CHECK5-NEXT:    store i8 [[CALL]], i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK5-NEXT:    [[TMP2:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
-// CHECK5-NEXT:    [[TMP3:%.*]] = zext i8 [[TMP2]] to i32
+// CHECK5-NEXT:    [[TMP10:%.*]] = load i8, i8* [[DOTCAPTURE_EXPR_]], align 1
+// CHECK5-NEXT:    [[TMP11:%.*]] = zext i8 [[TMP10]] to i32
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 100)
-// CHECK5-NEXT:    [[TMP4:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l40.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 [[TMP3]])
-// CHECK5-NEXT:    [[TMP5:%.*]] = icmp ne i32 [[TMP4]], 0
-// CHECK5-NEXT:    br i1 [[TMP5]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK5:       omp_offload.failed2:
+// CHECK5-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP12]], align 4
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 0, i32* [[TMP13]], align 4
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK5-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* null, i64** [[TMP16]], align 8
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* null, i64** [[TMP17]], align 8
+// CHECK5-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP18]], align 8
+// CHECK5-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP19]], align 8
+// CHECK5-NEXT:    [[TMP20:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 [[TMP11]], i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l40.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK5-NEXT:    [[TMP21:%.*]] = icmp ne i32 [[TMP20]], 0
+// CHECK5-NEXT:    br i1 [[TMP21]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK5:       omp_offload.failed3:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainI1SLi1EEiv_l40() #[[ATTR6]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK5:       omp_offload.cont3:
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK5:       omp_offload.cont4:
 // CHECK5-NEXT:    ret i32 0
 // CHECK5:       terminate.lpad:
-// CHECK5-NEXT:    [[TMP6:%.*]] = landingpad { i8*, i32 }
+// CHECK5-NEXT:    [[TMP22:%.*]] = landingpad { i8*, i32 }
 // CHECK5-NEXT:    catch i8* null
-// CHECK5-NEXT:    [[TMP7:%.*]] = extractvalue { i8*, i32 } [[TMP6]], 0
-// CHECK5-NEXT:    call void @__clang_call_terminate(i8* [[TMP7]]) #[[ATTR9]]
+// CHECK5-NEXT:    [[TMP23:%.*]] = extractvalue { i8*, i32 } [[TMP22]], 0
+// CHECK5-NEXT:    call void @__clang_call_terminate(i8* [[TMP23]]) #[[ATTR9]]
 // CHECK5-NEXT:    unreachable
 //
 //

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_simd_private_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_simd_private_codegen.cpp
index c88c9c8..66b0d84 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_simd_private_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_simd_private_codegen.cpp
@@ -264,9 +264,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -525,27 +542,44 @@ int main() {
 // CHECK1-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 8
 // CHECK1-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 8
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK1-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP4]]
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev
@@ -985,9 +1019,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l98() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1240,27 +1291,44 @@ int main() {
 // CHECK3-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 4
 // CHECK3-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK3-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP4]]
+// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_simd_proc_bind_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_simd_proc_bind_codegen.cpp
index 3909015..73f800c 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_simd_proc_bind_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_simd_proc_bind_codegen.cpp
@@ -60,21 +60,55 @@ int main() {
 // CHECK1-NEXT:    [[_TMP1:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l36.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l36() #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP2:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l39.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP3:%.*]] = icmp ne i32 [[TMP2]], 0
-// CHECK1-NEXT:    br i1 [[TMP3]], label [[OMP_OFFLOAD_FAILED2:%.*]], label [[OMP_OFFLOAD_CONT3:%.*]]
-// CHECK1:       omp_offload.failed2:
+// CHECK1-NEXT:    [[KERNEL_ARGS2:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l39.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS2]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED3:%.*]], label [[OMP_OFFLOAD_CONT4:%.*]]
+// CHECK1:       omp_offload.failed3:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l39() #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT3]]
-// CHECK1:       omp_offload.cont3:
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT4]]
+// CHECK1:       omp_offload.cont4:
 // CHECK1-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK1-NEXT:    ret i32 [[CALL]]
 //
@@ -388,9 +422,26 @@ int main() {
 // CHECK1-NEXT:  entry:
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 1000)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l29.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l29.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l29() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_simd_reduction_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_simd_reduction_codegen.cpp
index 80b52e8..5aca049 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_simd_reduction_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_simd_reduction_codegen.cpp
@@ -122,9 +122,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l70.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l70.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l70(i64 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -421,9 +438,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP10:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[TMP10]], 0
-// CHECK1-NEXT:    br i1 [[TMP11]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP9]], i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i64 [[TMP2]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -722,9 +756,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l70.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l70.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK3-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l70(i32 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1015,9 +1066,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP10:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB4]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[TMP10]], 0
-// CHECK3-NEXT:    br i1 [[TMP11]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP9]], i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.7, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.8, i32 0, i32 0), i64** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB4]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK3-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i32 [[TMP2]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_simd_schedule_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_simd_schedule_codegen.cpp
index b21216a..fe3cfc7 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_parallel_for_simd_schedule_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_simd_schedule_codegen.cpp
@@ -238,18 +238,18 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -264,93 +264,178 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l41.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l41.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l41(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l46.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l46.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK1:       omp_offload.failed16:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l46(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK1:       omp_offload.cont15:
-// CHECK1-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 8
-// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 8
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP31]], align 8
-// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK1:       omp_offload.cont17:
+// CHECK1-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 8
+// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 8
+// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP55]], align 8
+// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l52.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK1-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK1:       omp_offload.failed21:
+// CHECK1-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK1-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK1-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 8
+// CHECK1-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 8
+// CHECK1-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 8
+// CHECK1-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 8
+// CHECK1-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP64]], align 8
+// CHECK1-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP65]], align 8
+// CHECK1-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK1-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK1-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK1:       omp_offload.failed24:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l52(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK1:       omp_offload.cont22:
-// CHECK1-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 8
-// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 8
-// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK1:       omp_offload.cont25:
+// CHECK1-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 8
+// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP72]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l58.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK1-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK1:       omp_offload.failed28:
+// CHECK1-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK1-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 8
+// CHECK1-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 8
+// CHECK1-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 8
+// CHECK1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 8
+// CHECK1-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP81]], align 8
+// CHECK1-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP82]], align 8
+// CHECK1-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l58.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK1-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK1-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK1:       omp_offload.failed32:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l58(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK1:       omp_offload.cont29:
-// CHECK1-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP45]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK1:       omp_offload.cont33:
+// CHECK1-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i64 0, i64 0
+// CHECK1-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36
@@ -1222,18 +1307,18 @@ int main (int argc, char **argv) {
 // CHECK2-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK2-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK2-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK2-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 8
-// CHECK2-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 8
-// CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 8
-// CHECK2-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 8
-// CHECK2-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 8
-// CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 8
-// CHECK2-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK2-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK2-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 8
+// CHECK2-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 8
+// CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 8
+// CHECK2-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK2-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 8
+// CHECK2-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 8
+// CHECK2-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 8
+// CHECK2-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK2-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK2-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK2-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -1248,93 +1333,178 @@ int main (int argc, char **argv) {
 // CHECK2-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK2-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK2-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK2-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK2-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK2-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK2-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK2-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK2-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK2-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK2-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK2-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK2-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK2-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK2-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK2-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK2-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK2-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK2-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK2-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK2-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK2:       omp_offload.failed:
 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK2:       omp_offload.cont:
 // CHECK2-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK2-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK2-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK2-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK2-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK2-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK2-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK2-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK2-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK2-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK2-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK2-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK2-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK2-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l41.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK2-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK2-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK2:       omp_offload.failed7:
+// CHECK2-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK2-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK2-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK2-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK2-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK2-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK2-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK2-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK2-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK2-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK2-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK2-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK2-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK2-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l41.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK2-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK2-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK2:       omp_offload.failed8:
 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l41(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK2:       omp_offload.cont8:
-// CHECK2-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK2-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK2-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK2-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK2-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK2-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK2-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK2:       omp_offload.cont9:
+// CHECK2-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK2-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK2-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK2-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK2-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK2-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK2-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK2-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l46.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK2-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK2-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK2:       omp_offload.failed14:
+// CHECK2-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK2-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK2-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK2-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK2-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK2-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK2-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK2-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK2-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK2-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK2-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK2-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK2-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK2-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l46.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK2-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK2-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK2:       omp_offload.failed16:
 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l46(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK2:       omp_offload.cont15:
-// CHECK2-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK2-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 8
-// CHECK2-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK2-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 8
-// CHECK2-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i64 0, i64 0
-// CHECK2-NEXT:    store i8* null, i8** [[TMP31]], align 8
-// CHECK2-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK2:       omp_offload.cont17:
+// CHECK2-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK2-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 8
+// CHECK2-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK2-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 8
+// CHECK2-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i64 0, i64 0
+// CHECK2-NEXT:    store i8* null, i8** [[TMP55]], align 8
+// CHECK2-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK2-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l52.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK2-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK2-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK2:       omp_offload.failed21:
+// CHECK2-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK2-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK2-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK2-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK2-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK2-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK2-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 8
+// CHECK2-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK2-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 8
+// CHECK2-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 8
+// CHECK2-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 8
+// CHECK2-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP64]], align 8
+// CHECK2-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP65]], align 8
+// CHECK2-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK2-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK2-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK2:       omp_offload.failed24:
 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l52(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK2:       omp_offload.cont22:
-// CHECK2-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK2-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 8
-// CHECK2-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK2-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 8
-// CHECK2-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 0
-// CHECK2-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK2-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK2-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK2:       omp_offload.cont25:
+// CHECK2-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK2-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 8
+// CHECK2-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK2-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 8
+// CHECK2-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i64 0, i64 0
+// CHECK2-NEXT:    store i8* null, i8** [[TMP72]], align 8
+// CHECK2-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK2-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK2-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK2-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l58.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK2-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK2-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK2:       omp_offload.failed28:
+// CHECK2-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK2-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK2-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK2-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK2-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK2-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK2-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 8
+// CHECK2-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK2-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 8
+// CHECK2-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 8
+// CHECK2-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK2-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 8
+// CHECK2-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP81]], align 8
+// CHECK2-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK2-NEXT:    store i8** null, i8*** [[TMP82]], align 8
+// CHECK2-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l58.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK2-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK2-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK2:       omp_offload.failed32:
 // CHECK2-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l58(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK2:       omp_offload.cont29:
-// CHECK2-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK2-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i64 0, i64 0
-// CHECK2-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK2-NEXT:    ret i32 [[TMP45]]
+// CHECK2-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK2:       omp_offload.cont33:
+// CHECK2-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK2-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i64 0, i64 0
+// CHECK2-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK2-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36
@@ -2206,18 +2376,18 @@ int main (int argc, char **argv) {
 // CHECK5-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK5-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK5-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 4
-// CHECK5-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 4
-// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 4
-// CHECK5-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK5-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 4
+// CHECK5-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 4
+// CHECK5-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 4
+// CHECK5-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK5-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK5-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK5-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -2232,93 +2402,178 @@ int main (int argc, char **argv) {
 // CHECK5-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK5-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK5-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK5-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK5-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK5-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK5-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK5-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK5-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK5-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK5-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK5-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK5:       omp_offload.failed:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK5:       omp_offload.cont:
 // CHECK5-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK5-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK5-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK5-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK5-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK5-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK5-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l41.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK5-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK5:       omp_offload.failed7:
+// CHECK5-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK5-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK5-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK5-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK5-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK5-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l41.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK5-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK5-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK5:       omp_offload.failed8:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l41(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK5:       omp_offload.cont8:
-// CHECK5-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK5-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK5-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK5-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK5:       omp_offload.cont9:
+// CHECK5-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK5-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK5-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l46.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK5-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK5:       omp_offload.failed14:
+// CHECK5-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK5-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK5-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK5-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK5-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK5-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK5-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK5-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK5-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l46.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK5-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK5-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK5:       omp_offload.failed16:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l46(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK5:       omp_offload.cont15:
-// CHECK5-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 4
-// CHECK5-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 4
-// CHECK5-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP31]], align 4
-// CHECK5-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK5:       omp_offload.cont17:
+// CHECK5-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 4
+// CHECK5-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 4
+// CHECK5-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP55]], align 4
+// CHECK5-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l52.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK5-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK5:       omp_offload.failed21:
+// CHECK5-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK5-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK5-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 4
+// CHECK5-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 4
+// CHECK5-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 4
+// CHECK5-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 4
+// CHECK5-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP64]], align 4
+// CHECK5-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP65]], align 4
+// CHECK5-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK5-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK5-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK5:       omp_offload.failed24:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l52(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK5:       omp_offload.cont22:
-// CHECK5-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 4
-// CHECK5-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK5-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 4
-// CHECK5-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
-// CHECK5-NEXT:    store i8* null, i8** [[TMP40]], align 4
-// CHECK5-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK5:       omp_offload.cont25:
+// CHECK5-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK5-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 4
+// CHECK5-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK5-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 4
+// CHECK5-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i32 0, i32 0
+// CHECK5-NEXT:    store i8* null, i8** [[TMP72]], align 4
+// CHECK5-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK5-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK5-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l58.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK5-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK5-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK5:       omp_offload.failed28:
+// CHECK5-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK5-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK5-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK5-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK5-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK5-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK5-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 4
+// CHECK5-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK5-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 4
+// CHECK5-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 4
+// CHECK5-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK5-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 4
+// CHECK5-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP81]], align 4
+// CHECK5-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK5-NEXT:    store i8** null, i8*** [[TMP82]], align 4
+// CHECK5-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l58.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK5-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK5-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK5:       omp_offload.failed32:
 // CHECK5-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l58(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK5:       omp_offload.cont29:
-// CHECK5-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i32 0, i32 0
-// CHECK5-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK5-NEXT:    ret i32 [[TMP45]]
+// CHECK5-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK5:       omp_offload.cont33:
+// CHECK5-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK5-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i32 0, i32 0
+// CHECK5-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK5-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36
@@ -3163,18 +3418,18 @@ int main (int argc, char **argv) {
 // CHECK6-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK6-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK6-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK6-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK6-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK6-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK6-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
-// CHECK6-NEXT:    [[DOTOFFLOAD_BASEPTRS17:%.*]] = alloca [1 x i8*], align 4
-// CHECK6-NEXT:    [[DOTOFFLOAD_PTRS18:%.*]] = alloca [1 x i8*], align 4
-// CHECK6-NEXT:    [[DOTOFFLOAD_MAPPERS19:%.*]] = alloca [1 x i8*], align 4
-// CHECK6-NEXT:    [[_TMP20:%.*]] = alloca i32, align 4
-// CHECK6-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [1 x i8*], align 4
-// CHECK6-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [1 x i8*], align 4
-// CHECK6-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [1 x i8*], align 4
-// CHECK6-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
+// CHECK6-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK6-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK6-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK6-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
+// CHECK6-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [1 x i8*], align 4
+// CHECK6-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [1 x i8*], align 4
+// CHECK6-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [1 x i8*], align 4
+// CHECK6-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
+// CHECK6-NEXT:    [[DOTOFFLOAD_BASEPTRS27:%.*]] = alloca [1 x i8*], align 4
+// CHECK6-NEXT:    [[DOTOFFLOAD_PTRS28:%.*]] = alloca [1 x i8*], align 4
+// CHECK6-NEXT:    [[DOTOFFLOAD_MAPPERS29:%.*]] = alloca [1 x i8*], align 4
+// CHECK6-NEXT:    [[_TMP30:%.*]] = alloca i32, align 4
 // CHECK6-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK6-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK6-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -3189,93 +3444,178 @@ int main (int argc, char **argv) {
 // CHECK6-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK6-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK6-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 123)
-// CHECK6-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK6-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK6-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK6-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK6-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK6-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK6-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK6-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK6-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK6-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK6-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK6-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK6-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK6-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK6-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK6-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK6-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK6-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK6-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK6-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK6-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK6-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK6-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK6:       omp_offload.failed:
 // CHECK6-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK6:       omp_offload.cont:
 // CHECK6-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK6-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK6-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK6-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK6-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK6-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK6-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK6-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK6-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK6-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK6-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK6-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK6-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK6-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK6-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l41.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK6-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK6-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK6:       omp_offload.failed7:
+// CHECK6-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK6-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK6-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK6-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK6-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK6-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK6-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK6-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK6-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK6-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK6-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK6-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK6-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK6-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK6-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK6-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK6-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK6-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l41.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK6-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK6-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK6:       omp_offload.failed8:
 // CHECK6-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l41(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK6:       omp_offload.cont8:
-// CHECK6-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK6-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK6-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK6-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK6-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK6-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK6-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK6:       omp_offload.cont9:
+// CHECK6-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK6-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK6-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK6-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK6-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK6-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK6-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK6-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK6-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l46.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK6-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK6-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK6:       omp_offload.failed14:
+// CHECK6-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK6-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK6-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK6-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK6-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK6-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK6-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK6-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK6-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK6-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK6-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK6-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK6-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK6-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK6-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK6-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK6-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK6-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l46.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK6-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK6-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK6:       omp_offload.failed16:
 // CHECK6-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l46(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK6:       omp_offload.cont15:
-// CHECK6-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to %struct.SS**
-// CHECK6-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP28]], align 4
-// CHECK6-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP30:%.*]] = bitcast i8** [[TMP29]] to [123 x i32]**
-// CHECK6-NEXT:    store [123 x i32]* [[A16]], [123 x i32]** [[TMP30]], align 4
-// CHECK6-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS19]], i32 0, i32 0
-// CHECK6-NEXT:    store i8* null, i8** [[TMP31]], align 4
-// CHECK6-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS17]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS18]], i32 0, i32 0
+// CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK6:       omp_offload.cont17:
+// CHECK6-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to %struct.SS**
+// CHECK6-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP52]], align 4
+// CHECK6-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to [123 x i32]**
+// CHECK6-NEXT:    store [123 x i32]* [[A18]], [123 x i32]** [[TMP54]], align 4
+// CHECK6-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
+// CHECK6-NEXT:    store i8* null, i8** [[TMP55]], align 4
+// CHECK6-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
 // CHECK6-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK6-NEXT:    [[TMP34:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l52.region_id, i32 1, i8** [[TMP32]], i8** [[TMP33]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK6-NEXT:    [[TMP35:%.*]] = icmp ne i32 [[TMP34]], 0
-// CHECK6-NEXT:    br i1 [[TMP35]], label [[OMP_OFFLOAD_FAILED21:%.*]], label [[OMP_OFFLOAD_CONT22:%.*]]
-// CHECK6:       omp_offload.failed21:
+// CHECK6-NEXT:    [[KERNEL_ARGS23:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK6-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 0
+// CHECK6-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK6-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 1
+// CHECK6-NEXT:    store i32 1, i32* [[TMP59]], align 4
+// CHECK6-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 2
+// CHECK6-NEXT:    store i8** [[TMP56]], i8*** [[TMP60]], align 4
+// CHECK6-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 3
+// CHECK6-NEXT:    store i8** [[TMP57]], i8*** [[TMP61]], align 4
+// CHECK6-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 4
+// CHECK6-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.12, i32 0, i32 0), i64** [[TMP62]], align 4
+// CHECK6-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 5
+// CHECK6-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP63]], align 4
+// CHECK6-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 6
+// CHECK6-NEXT:    store i8** null, i8*** [[TMP64]], align 4
+// CHECK6-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]], i32 0, i32 7
+// CHECK6-NEXT:    store i8** null, i8*** [[TMP65]], align 4
+// CHECK6-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l52.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS23]])
+// CHECK6-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK6-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
+// CHECK6:       omp_offload.failed24:
 // CHECK6-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l52(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT22]]
-// CHECK6:       omp_offload.cont22:
-// CHECK6-NEXT:    [[A23:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to %struct.SS**
-// CHECK6-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP37]], align 4
-// CHECK6-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to [123 x i32]**
-// CHECK6-NEXT:    store [123 x i32]* [[A23]], [123 x i32]** [[TMP39]], align 4
-// CHECK6-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i32 0, i32 0
-// CHECK6-NEXT:    store i8* null, i8** [[TMP40]], align 4
-// CHECK6-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
+// CHECK6:       omp_offload.cont25:
+// CHECK6-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP69:%.*]] = bitcast i8** [[TMP68]] to %struct.SS**
+// CHECK6-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP69]], align 4
+// CHECK6-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to [123 x i32]**
+// CHECK6-NEXT:    store [123 x i32]* [[A26]], [123 x i32]** [[TMP71]], align 4
+// CHECK6-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS29]], i32 0, i32 0
+// CHECK6-NEXT:    store i8* null, i8** [[TMP72]], align 4
+// CHECK6-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS27]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS28]], i32 0, i32 0
 // CHECK6-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 123)
-// CHECK6-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l58.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK6-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK6-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
-// CHECK6:       omp_offload.failed28:
+// CHECK6-NEXT:    [[KERNEL_ARGS31:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK6-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 0
+// CHECK6-NEXT:    store i32 1, i32* [[TMP75]], align 4
+// CHECK6-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 1
+// CHECK6-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK6-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 2
+// CHECK6-NEXT:    store i8** [[TMP73]], i8*** [[TMP77]], align 4
+// CHECK6-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 3
+// CHECK6-NEXT:    store i8** [[TMP74]], i8*** [[TMP78]], align 4
+// CHECK6-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 4
+// CHECK6-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.16, i32 0, i32 0), i64** [[TMP79]], align 4
+// CHECK6-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 5
+// CHECK6-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP80]], align 4
+// CHECK6-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 6
+// CHECK6-NEXT:    store i8** null, i8*** [[TMP81]], align 4
+// CHECK6-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]], i32 0, i32 7
+// CHECK6-NEXT:    store i8** null, i8*** [[TMP82]], align 4
+// CHECK6-NEXT:    [[TMP83:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l58.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS31]])
+// CHECK6-NEXT:    [[TMP84:%.*]] = icmp ne i32 [[TMP83]], 0
+// CHECK6-NEXT:    br i1 [[TMP84]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
+// CHECK6:       omp_offload.failed32:
 // CHECK6-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l58(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
-// CHECK6:       omp_offload.cont29:
-// CHECK6-NEXT:    [[A30:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK6-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A30]], i32 0, i32 0
-// CHECK6-NEXT:    [[TMP45:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK6-NEXT:    ret i32 [[TMP45]]
+// CHECK6-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
+// CHECK6:       omp_offload.cont33:
+// CHECK6-NEXT:    [[A34:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK6-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A34]], i32 0, i32 0
+// CHECK6-NEXT:    [[TMP85:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK6-NEXT:    ret i32 [[TMP85]]
 //
 //
 // CHECK6-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l36
@@ -4512,31 +4852,31 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[M_CASTED:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES24:%.*]] = alloca [4 x i64], align 8
-// CHECK13-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
+// CHECK13-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[N_CASTED34:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS36:%.*]] = alloca [3 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS37:%.*]] = alloca [3 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS38:%.*]] = alloca [3 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES39:%.*]] = alloca [3 x i64], align 8
-// CHECK13-NEXT:    [[_TMP40:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_41:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_42:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[M_CASTED49:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[N_CASTED51:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS53:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS54:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS55:%.*]] = alloca [4 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES56:%.*]] = alloca [4 x i64], align 8
-// CHECK13-NEXT:    [[_TMP57:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_58:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_59:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[N_CASTED36:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS38:%.*]] = alloca [3 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS39:%.*]] = alloca [3 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS40:%.*]] = alloca [3 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES41:%.*]] = alloca [3 x i64], align 8
+// CHECK13-NEXT:    [[_TMP42:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_43:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_44:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[M_CASTED52:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[N_CASTED54:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS56:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS57:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS58:%.*]] = alloca [4 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_SIZES59:%.*]] = alloca [4 x i64], align 8
+// CHECK13-NEXT:    [[_TMP60:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_61:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTCAPTURE_EXPR_62:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK13-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK13-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -4595,260 +4935,345 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK13-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK13-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK13-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK13-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK13-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK13-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK13-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK13-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK13-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK13-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK13-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK13-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       omp_offload.cont:
-// CHECK13-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK13-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK13-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK13-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
-// CHECK13-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK13-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK13-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK13-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK13-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK13-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK13-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK13-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK13-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK13-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK13-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK13-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK13-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK13-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK13-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK13-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
+// CHECK13-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK13-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK13-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK13-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK13-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK13-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK13-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK13-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK13-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK13-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK13-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK13-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK13-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK13-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK13-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK13-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK13-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK13-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP58]])
-// CHECK13-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK13-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK13:       omp_offload.failed16:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK13:       omp_offload.cont17:
-// CHECK13-NEXT:    [[TMP61:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    [[CONV18:%.*]] = bitcast i64* [[M_CASTED]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP61]], i32* [[CONV18]], align 4
-// CHECK13-NEXT:    [[TMP62:%.*]] = load i64, i64* [[M_CASTED]], align 8
-// CHECK13-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP63]], i32* [[CONV20]], align 4
-// CHECK13-NEXT:    [[TMP64:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK13-NEXT:    [[TMP65:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP66:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES24]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP66]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
-// CHECK13-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP62]], i64* [[TMP68]], align 8
-// CHECK13-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP62]], i64* [[TMP70]], align 8
-// CHECK13-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP71]], align 8
-// CHECK13-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP64]], i64* [[TMP73]], align 8
-// CHECK13-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP64]], i64* [[TMP75]], align 8
-// CHECK13-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP76]], align 8
-// CHECK13-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP78]], align 8
-// CHECK13-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP80]], align 8
-// CHECK13-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP81]], align 8
-// CHECK13-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
-// CHECK13-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 8
-// CHECK13-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
-// CHECK13-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP85]], align 8
-// CHECK13-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 3
-// CHECK13-NEXT:    store i64 [[TMP65]], i64* [[TMP86]], align 8
-// CHECK13-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK13-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK13-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK13-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP66]])
+// CHECK13-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK13-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK13-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK13-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK13-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK13-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK13-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK13-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK13-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK13-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK13-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK13:       omp_offload.failed17:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK13:       omp_offload.cont18:
+// CHECK13-NEXT:    [[TMP77:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    [[CONV19:%.*]] = bitcast i64* [[M_CASTED]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP77]], i32* [[CONV19]], align 4
+// CHECK13-NEXT:    [[TMP78:%.*]] = load i64, i64* [[M_CASTED]], align 8
+// CHECK13-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP79]], i32* [[CONV21]], align 4
+// CHECK13-NEXT:    [[TMP80:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK13-NEXT:    [[TMP81:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP82:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP82]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
+// CHECK13-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP78]], i64* [[TMP84]], align 8
+// CHECK13-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP78]], i64* [[TMP86]], align 8
+// CHECK13-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
 // CHECK13-NEXT:    store i8* null, i8** [[TMP87]], align 8
-// CHECK13-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP91:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP91]], i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK13-NEXT:    [[TMP92:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK13-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[TMP92]], 0
-// CHECK13-NEXT:    [[DIV29:%.*]] = sdiv i32 [[SUB28]], 1
-// CHECK13-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[DIV29]], 1
-// CHECK13-NEXT:    store i32 [[SUB30]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK13-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK13-NEXT:    [[ADD31:%.*]] = add nsw i32 [[TMP93]], 1
-// CHECK13-NEXT:    [[TMP94:%.*]] = zext i32 [[ADD31]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP94]])
-// CHECK13-NEXT:    [[TMP95:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164.region_id, i32 4, i8** [[TMP88]], i8** [[TMP89]], i64* [[TMP90]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP96:%.*]] = icmp ne i32 [[TMP95]], 0
-// CHECK13-NEXT:    br i1 [[TMP96]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
-// CHECK13:       omp_offload.failed32:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164(i64 [[TMP62]], i64 [[TMP64]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
-// CHECK13:       omp_offload.cont33:
-// CHECK13-NEXT:    [[TMP97:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    [[CONV35:%.*]] = bitcast i64* [[N_CASTED34]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP97]], i32* [[CONV35]], align 4
-// CHECK13-NEXT:    [[TMP98:%.*]] = load i64, i64* [[N_CASTED34]], align 8
-// CHECK13-NEXT:    [[TMP99:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP100:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES39]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP100]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
-// CHECK13-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP98]], i64* [[TMP102]], align 8
-// CHECK13-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP98]], i64* [[TMP104]], align 8
-// CHECK13-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP105]], align 8
-// CHECK13-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP107]], align 8
-// CHECK13-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP109]], align 8
-// CHECK13-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP110]], align 8
-// CHECK13-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP112]], align 8
-// CHECK13-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP114]], align 8
-// CHECK13-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES39]], i32 0, i32 2
-// CHECK13-NEXT:    store i64 [[TMP99]], i64* [[TMP115]], align 8
-// CHECK13-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP116]], align 8
-// CHECK13-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES39]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP120:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP120]], i32* [[DOTCAPTURE_EXPR_41]], align 4
-// CHECK13-NEXT:    [[TMP121:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_41]], align 4
-// CHECK13-NEXT:    [[SUB43:%.*]] = sub nsw i32 [[TMP121]], 0
-// CHECK13-NEXT:    [[DIV44:%.*]] = sdiv i32 [[SUB43]], 1
-// CHECK13-NEXT:    [[SUB45:%.*]] = sub nsw i32 [[DIV44]], 1
-// CHECK13-NEXT:    store i32 [[SUB45]], i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK13-NEXT:    [[TMP122:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK13-NEXT:    [[ADD46:%.*]] = add nsw i32 [[TMP122]], 1
-// CHECK13-NEXT:    [[TMP123:%.*]] = zext i32 [[ADD46]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP123]])
-// CHECK13-NEXT:    [[TMP124:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169.region_id, i32 3, i8** [[TMP117]], i8** [[TMP118]], i64* [[TMP119]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP125:%.*]] = icmp ne i32 [[TMP124]], 0
-// CHECK13-NEXT:    br i1 [[TMP125]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
-// CHECK13:       omp_offload.failed47:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169(i64 [[TMP98]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
-// CHECK13:       omp_offload.cont48:
-// CHECK13-NEXT:    [[TMP126:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    [[CONV50:%.*]] = bitcast i64* [[M_CASTED49]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP126]], i32* [[CONV50]], align 4
-// CHECK13-NEXT:    [[TMP127:%.*]] = load i64, i64* [[M_CASTED49]], align 8
-// CHECK13-NEXT:    [[TMP128:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    [[CONV52:%.*]] = bitcast i64* [[N_CASTED51]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP128]], i32* [[CONV52]], align 4
-// CHECK13-NEXT:    [[TMP129:%.*]] = load i64, i64* [[N_CASTED51]], align 8
-// CHECK13-NEXT:    [[TMP130:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK13-NEXT:    [[TMP131:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES56]] to i8*
-// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP131]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
-// CHECK13-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP80]], i64* [[TMP89]], align 8
+// CHECK13-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP80]], i64* [[TMP91]], align 8
+// CHECK13-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP92]], align 8
+// CHECK13-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP94]], align 8
+// CHECK13-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP96]], align 8
+// CHECK13-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP97]], align 8
+// CHECK13-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 8
+// CHECK13-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP101]], align 8
+// CHECK13-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 3
+// CHECK13-NEXT:    store i64 [[TMP81]], i64* [[TMP102]], align 8
+// CHECK13-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK13-NEXT:    store i8* null, i8** [[TMP103]], align 8
+// CHECK13-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP107:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP107]], i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK13-NEXT:    [[TMP108:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK13-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP108]], 0
+// CHECK13-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
+// CHECK13-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
+// CHECK13-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK13-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK13-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP109]], 1
+// CHECK13-NEXT:    [[TMP110:%.*]] = zext i32 [[ADD32]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP110]])
+// CHECK13-NEXT:    [[KERNEL_ARGS33:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP111]], align 4
+// CHECK13-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 4, i32* [[TMP112]], align 4
+// CHECK13-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP104]], i8*** [[TMP113]], align 8
+// CHECK13-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK13-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP106]], i64** [[TMP115]], align 8
+// CHECK13-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP116]], align 8
+// CHECK13-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP117]], align 8
+// CHECK13-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK13-NEXT:    [[TMP119:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]])
+// CHECK13-NEXT:    [[TMP120:%.*]] = icmp ne i32 [[TMP119]], 0
+// CHECK13-NEXT:    br i1 [[TMP120]], label [[OMP_OFFLOAD_FAILED34:%.*]], label [[OMP_OFFLOAD_CONT35:%.*]]
+// CHECK13:       omp_offload.failed34:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164(i64 [[TMP78]], i64 [[TMP80]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT35]]
+// CHECK13:       omp_offload.cont35:
+// CHECK13-NEXT:    [[TMP121:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[CONV37:%.*]] = bitcast i64* [[N_CASTED36]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP121]], i32* [[CONV37]], align 4
+// CHECK13-NEXT:    [[TMP122:%.*]] = load i64, i64* [[N_CASTED36]], align 8
+// CHECK13-NEXT:    [[TMP123:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP124:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES41]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP124]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
+// CHECK13-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP122]], i64* [[TMP126]], align 8
+// CHECK13-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP122]], i64* [[TMP128]], align 8
+// CHECK13-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS40]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP129]], align 8
+// CHECK13-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP131]], align 8
+// CHECK13-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 1
 // CHECK13-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP127]], i64* [[TMP133]], align 8
-// CHECK13-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP127]], i64* [[TMP135]], align 8
-// CHECK13-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP136]], align 8
-// CHECK13-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP129]], i64* [[TMP138]], align 8
-// CHECK13-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP129]], i64* [[TMP140]], align 8
-// CHECK13-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP141]], align 8
-// CHECK13-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP143]], align 8
-// CHECK13-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 2
-// CHECK13-NEXT:    [[TMP145:%.*]] = bitcast i8** [[TMP144]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP145]], align 8
-// CHECK13-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 2
-// CHECK13-NEXT:    store i8* null, i8** [[TMP146]], align 8
-// CHECK13-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 3
-// CHECK13-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP148]], align 8
-// CHECK13-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 3
-// CHECK13-NEXT:    [[TMP150:%.*]] = bitcast i8** [[TMP149]] to i32**
-// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP150]], align 8
-// CHECK13-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES56]], i32 0, i32 3
-// CHECK13-NEXT:    store i64 [[TMP130]], i64* [[TMP151]], align 8
-// CHECK13-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 3
-// CHECK13-NEXT:    store i8* null, i8** [[TMP152]], align 8
-// CHECK13-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES56]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP156:%.*]] = load i32, i32* [[N]], align 4
-// CHECK13-NEXT:    store i32 [[TMP156]], i32* [[DOTCAPTURE_EXPR_58]], align 4
-// CHECK13-NEXT:    [[TMP157:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_58]], align 4
-// CHECK13-NEXT:    [[SUB60:%.*]] = sub nsw i32 [[TMP157]], 0
-// CHECK13-NEXT:    [[DIV61:%.*]] = sdiv i32 [[SUB60]], 1
-// CHECK13-NEXT:    [[SUB62:%.*]] = sub nsw i32 [[DIV61]], 1
-// CHECK13-NEXT:    store i32 [[SUB62]], i32* [[DOTCAPTURE_EXPR_59]], align 4
-// CHECK13-NEXT:    [[TMP158:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_59]], align 4
-// CHECK13-NEXT:    [[ADD63:%.*]] = add nsw i32 [[TMP158]], 1
-// CHECK13-NEXT:    [[TMP159:%.*]] = zext i32 [[ADD63]] to i64
-// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP159]])
-// CHECK13-NEXT:    [[TMP160:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174.region_id, i32 4, i8** [[TMP153]], i8** [[TMP154]], i64* [[TMP155]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP161:%.*]] = icmp ne i32 [[TMP160]], 0
-// CHECK13-NEXT:    br i1 [[TMP161]], label [[OMP_OFFLOAD_FAILED64:%.*]], label [[OMP_OFFLOAD_CONT65:%.*]]
-// CHECK13:       omp_offload.failed64:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174(i64 [[TMP127]], i64 [[TMP129]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT65]]
-// CHECK13:       omp_offload.cont65:
-// CHECK13-NEXT:    [[TMP162:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK13-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP162]])
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP133]], align 8
+// CHECK13-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS40]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP134]], align 8
+// CHECK13-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP136]], align 8
+// CHECK13-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP138]], align 8
+// CHECK13-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES41]], i32 0, i32 2
+// CHECK13-NEXT:    store i64 [[TMP123]], i64* [[TMP139]], align 8
+// CHECK13-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS40]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP140]], align 8
+// CHECK13-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES41]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP144:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP144]], i32* [[DOTCAPTURE_EXPR_43]], align 4
+// CHECK13-NEXT:    [[TMP145:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_43]], align 4
+// CHECK13-NEXT:    [[SUB45:%.*]] = sub nsw i32 [[TMP145]], 0
+// CHECK13-NEXT:    [[DIV46:%.*]] = sdiv i32 [[SUB45]], 1
+// CHECK13-NEXT:    [[SUB47:%.*]] = sub nsw i32 [[DIV46]], 1
+// CHECK13-NEXT:    store i32 [[SUB47]], i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK13-NEXT:    [[TMP146:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK13-NEXT:    [[ADD48:%.*]] = add nsw i32 [[TMP146]], 1
+// CHECK13-NEXT:    [[TMP147:%.*]] = zext i32 [[ADD48]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP147]])
+// CHECK13-NEXT:    [[KERNEL_ARGS49:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP148]], align 4
+// CHECK13-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 3, i32* [[TMP149]], align 4
+// CHECK13-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP141]], i8*** [[TMP150]], align 8
+// CHECK13-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP142]], i8*** [[TMP151]], align 8
+// CHECK13-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP143]], i64** [[TMP152]], align 8
+// CHECK13-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP153]], align 8
+// CHECK13-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP154]], align 8
+// CHECK13-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP155]], align 8
+// CHECK13-NEXT:    [[TMP156:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]])
+// CHECK13-NEXT:    [[TMP157:%.*]] = icmp ne i32 [[TMP156]], 0
+// CHECK13-NEXT:    br i1 [[TMP157]], label [[OMP_OFFLOAD_FAILED50:%.*]], label [[OMP_OFFLOAD_CONT51:%.*]]
+// CHECK13:       omp_offload.failed50:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169(i64 [[TMP122]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT51]]
+// CHECK13:       omp_offload.cont51:
+// CHECK13-NEXT:    [[TMP158:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    [[CONV53:%.*]] = bitcast i64* [[M_CASTED52]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP158]], i32* [[CONV53]], align 4
+// CHECK13-NEXT:    [[TMP159:%.*]] = load i64, i64* [[M_CASTED52]], align 8
+// CHECK13-NEXT:    [[TMP160:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    [[CONV55:%.*]] = bitcast i64* [[N_CASTED54]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP160]], i32* [[CONV55]], align 4
+// CHECK13-NEXT:    [[TMP161:%.*]] = load i64, i64* [[N_CASTED54]], align 8
+// CHECK13-NEXT:    [[TMP162:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK13-NEXT:    [[TMP163:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES59]] to i8*
+// CHECK13-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP163]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
+// CHECK13-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP165:%.*]] = bitcast i8** [[TMP164]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP159]], i64* [[TMP165]], align 8
+// CHECK13-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP159]], i64* [[TMP167]], align 8
+// CHECK13-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP168]], align 8
+// CHECK13-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP170:%.*]] = bitcast i8** [[TMP169]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP161]], i64* [[TMP170]], align 8
+// CHECK13-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP172:%.*]] = bitcast i8** [[TMP171]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP161]], i64* [[TMP172]], align 8
+// CHECK13-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP173]], align 8
+// CHECK13-NEXT:    [[TMP174:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP175:%.*]] = bitcast i8** [[TMP174]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP175]], align 8
+// CHECK13-NEXT:    [[TMP176:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 2
+// CHECK13-NEXT:    [[TMP177:%.*]] = bitcast i8** [[TMP176]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP1]], i64* [[TMP177]], align 8
+// CHECK13-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 2
+// CHECK13-NEXT:    store i8* null, i8** [[TMP178]], align 8
+// CHECK13-NEXT:    [[TMP179:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP180:%.*]] = bitcast i8** [[TMP179]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP180]], align 8
+// CHECK13-NEXT:    [[TMP181:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 3
+// CHECK13-NEXT:    [[TMP182:%.*]] = bitcast i8** [[TMP181]] to i32**
+// CHECK13-NEXT:    store i32* [[VLA]], i32** [[TMP182]], align 8
+// CHECK13-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES59]], i32 0, i32 3
+// CHECK13-NEXT:    store i64 [[TMP162]], i64* [[TMP183]], align 8
+// CHECK13-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 3
+// CHECK13-NEXT:    store i8* null, i8** [[TMP184]], align 8
+// CHECK13-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES59]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP188:%.*]] = load i32, i32* [[N]], align 4
+// CHECK13-NEXT:    store i32 [[TMP188]], i32* [[DOTCAPTURE_EXPR_61]], align 4
+// CHECK13-NEXT:    [[TMP189:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_61]], align 4
+// CHECK13-NEXT:    [[SUB63:%.*]] = sub nsw i32 [[TMP189]], 0
+// CHECK13-NEXT:    [[DIV64:%.*]] = sdiv i32 [[SUB63]], 1
+// CHECK13-NEXT:    [[SUB65:%.*]] = sub nsw i32 [[DIV64]], 1
+// CHECK13-NEXT:    store i32 [[SUB65]], i32* [[DOTCAPTURE_EXPR_62]], align 4
+// CHECK13-NEXT:    [[TMP190:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_62]], align 4
+// CHECK13-NEXT:    [[ADD66:%.*]] = add nsw i32 [[TMP190]], 1
+// CHECK13-NEXT:    [[TMP191:%.*]] = zext i32 [[ADD66]] to i64
+// CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP191]])
+// CHECK13-NEXT:    [[KERNEL_ARGS67:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP192]], align 4
+// CHECK13-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 4, i32* [[TMP193]], align 4
+// CHECK13-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP185]], i8*** [[TMP194]], align 8
+// CHECK13-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP186]], i8*** [[TMP195]], align 8
+// CHECK13-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* [[TMP187]], i64** [[TMP196]], align 8
+// CHECK13-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP197]], align 8
+// CHECK13-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP198]], align 8
+// CHECK13-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP199]], align 8
+// CHECK13-NEXT:    [[TMP200:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]])
+// CHECK13-NEXT:    [[TMP201:%.*]] = icmp ne i32 [[TMP200]], 0
+// CHECK13-NEXT:    br i1 [[TMP201]], label [[OMP_OFFLOAD_FAILED68:%.*]], label [[OMP_OFFLOAD_CONT69:%.*]]
+// CHECK13:       omp_offload.failed68:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174(i64 [[TMP159]], i64 [[TMP161]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT69]]
+// CHECK13:       omp_offload.cont69:
+// CHECK13-NEXT:    [[TMP202:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK13-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP202]])
 // CHECK13-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK13-NEXT:    [[TMP163:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK13-NEXT:    call void @llvm.stackrestore(i8* [[TMP163]])
-// CHECK13-NEXT:    [[TMP164:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK13-NEXT:    ret i32 [[TMP164]]
+// CHECK13-NEXT:    [[TMP203:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK13-NEXT:    call void @llvm.stackrestore(i8* [[TMP203]])
+// CHECK13-NEXT:    [[TMP204:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK13-NEXT:    ret i32 [[TMP204]]
 //
 //
 // CHECK13-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154
@@ -6108,19 +6533,19 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 8
 // CHECK13-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    [[M_CASTED:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 8
-// CHECK13-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK13-NEXT:    [[M_CASTED19:%.*]] = alloca i64, align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [2 x i8*], align 8
-// CHECK13-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 8
+// CHECK13-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK13-NEXT:    [[M_CASTED22:%.*]] = alloca i64, align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [2 x i8*], align 8
+// CHECK13-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK13-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK13-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK13-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -6134,109 +6559,194 @@ int main (int argc, char **argv) {
 // CHECK13-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK13-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l122.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK13-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK13-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK13-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK13-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK13-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK13-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK13-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK13-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK13-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l122.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK13-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK13-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK13:       omp_offload.failed:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l122([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK13:       omp_offload.cont:
-// CHECK13-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK13-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK13-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK13-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK13-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK13-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK13-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l127.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK13-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK13:       omp_offload.failed5:
+// CHECK13-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK13-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK13-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK13-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK13-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK13-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK13-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK13-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK13-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l127.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK13-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK13-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK13:       omp_offload.failed6:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l127([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK13:       omp_offload.cont6:
-// CHECK13-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK13:       omp_offload.cont7:
+// CHECK13-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
 // CHECK13-NEXT:    [[CONV:%.*]] = bitcast i64* [[M_CASTED]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP18]], i32* [[CONV]], align 4
-// CHECK13-NEXT:    [[TMP19:%.*]] = load i64, i64* [[M_CASTED]], align 8
-// CHECK13-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP19]], i64* [[TMP21]], align 8
-// CHECK13-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP19]], i64* [[TMP23]], align 8
-// CHECK13-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP24]], align 8
-// CHECK13-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP26]], align 8
-// CHECK13-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP28]], align 8
-// CHECK13-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP29]], align 8
-// CHECK13-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 [[TMP34]], i32* [[CONV]], align 4
+// CHECK13-NEXT:    [[TMP35:%.*]] = load i64, i64* [[M_CASTED]], align 8
+// CHECK13-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP35]], i64* [[TMP37]], align 8
+// CHECK13-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP35]], i64* [[TMP39]], align 8
+// CHECK13-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP40]], align 8
+// CHECK13-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP42]], align 8
+// CHECK13-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP44]], align 8
+// CHECK13-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP45]], align 8
+// CHECK13-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132.region_id, i32 2, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK13-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK13:       omp_offload.failed11:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132(i64 [[TMP19]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK13:       omp_offload.cont12:
-// CHECK13-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 8
-// CHECK13-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 8
-// CHECK13-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP38]], align 8
-// CHECK13-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK13-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK13-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 2, i32* [[TMP49]], align 4
+// CHECK13-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP46]], i8*** [[TMP50]], align 8
+// CHECK13-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 8
+// CHECK13-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK13-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK13-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK13-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK13-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK13-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK13:       omp_offload.failed13:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132(i64 [[TMP35]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK13:       omp_offload.cont14:
+// CHECK13-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP59]], align 8
+// CHECK13-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP61]], align 8
+// CHECK13-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP62]], align 8
+// CHECK13-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP41:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l137.region_id, i32 1, i8** [[TMP39]], i8** [[TMP40]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP42:%.*]] = icmp ne i32 [[TMP41]], 0
-// CHECK13-NEXT:    br i1 [[TMP42]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK13:       omp_offload.failed17:
+// CHECK13-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP65]], align 4
+// CHECK13-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK13-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP63]], i8*** [[TMP67]], align 8
+// CHECK13-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 8
+// CHECK13-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP69]], align 8
+// CHECK13-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP70]], align 8
+// CHECK13-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP71]], align 8
+// CHECK13-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP72]], align 8
+// CHECK13-NEXT:    [[TMP73:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l137.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK13-NEXT:    [[TMP74:%.*]] = icmp ne i32 [[TMP73]], 0
+// CHECK13-NEXT:    br i1 [[TMP74]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK13:       omp_offload.failed20:
 // CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l137([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK13:       omp_offload.cont18:
-// CHECK13-NEXT:    [[TMP43:%.*]] = load i32, i32* [[M]], align 4
-// CHECK13-NEXT:    [[CONV20:%.*]] = bitcast i64* [[M_CASTED19]] to i32*
-// CHECK13-NEXT:    store i32 [[TMP43]], i32* [[CONV20]], align 4
-// CHECK13-NEXT:    [[TMP44:%.*]] = load i64, i64* [[M_CASTED19]], align 8
-// CHECK13-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP44]], i64* [[TMP46]], align 8
-// CHECK13-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i64*
-// CHECK13-NEXT:    store i64 [[TMP44]], i64* [[TMP48]], align 8
-// CHECK13-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
-// CHECK13-NEXT:    store i8* null, i8** [[TMP49]], align 8
-// CHECK13-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP51]], align 8
-// CHECK13-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK13-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to [10 x i32]**
-// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP53]], align 8
-// CHECK13-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
-// CHECK13-NEXT:    store i8* null, i8** [[TMP54]], align 8
-// CHECK13-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK13-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK13:       omp_offload.cont21:
+// CHECK13-NEXT:    [[TMP75:%.*]] = load i32, i32* [[M]], align 4
+// CHECK13-NEXT:    [[CONV23:%.*]] = bitcast i64* [[M_CASTED22]] to i32*
+// CHECK13-NEXT:    store i32 [[TMP75]], i32* [[CONV23]], align 4
+// CHECK13-NEXT:    [[TMP76:%.*]] = load i64, i64* [[M_CASTED22]], align 8
+// CHECK13-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP76]], i64* [[TMP78]], align 8
+// CHECK13-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
+// CHECK13-NEXT:    store i64 [[TMP76]], i64* [[TMP80]], align 8
+// CHECK13-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 0
+// CHECK13-NEXT:    store i8* null, i8** [[TMP81]], align 8
+// CHECK13-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP83]], align 8
+// CHECK13-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 1
+// CHECK13-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to [10 x i32]**
+// CHECK13-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP85]], align 8
+// CHECK13-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 1
+// CHECK13-NEXT:    store i8* null, i8** [[TMP86]], align 8
+// CHECK13-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK13-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
 // CHECK13-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK13-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK13-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
-// CHECK13-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
-// CHECK13:       omp_offload.failed25:
-// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142(i64 [[TMP44]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
-// CHECK13:       omp_offload.cont26:
+// CHECK13-NEXT:    [[KERNEL_ARGS28:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK13-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 0
+// CHECK13-NEXT:    store i32 1, i32* [[TMP89]], align 4
+// CHECK13-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 1
+// CHECK13-NEXT:    store i32 2, i32* [[TMP90]], align 4
+// CHECK13-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 2
+// CHECK13-NEXT:    store i8** [[TMP87]], i8*** [[TMP91]], align 8
+// CHECK13-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 3
+// CHECK13-NEXT:    store i8** [[TMP88]], i8*** [[TMP92]], align 8
+// CHECK13-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 4
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP93]], align 8
+// CHECK13-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 5
+// CHECK13-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP94]], align 8
+// CHECK13-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 6
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP95]], align 8
+// CHECK13-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 7
+// CHECK13-NEXT:    store i8** null, i8*** [[TMP96]], align 8
+// CHECK13-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]])
+// CHECK13-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
+// CHECK13-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
+// CHECK13:       omp_offload.failed29:
+// CHECK13-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142(i64 [[TMP76]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK13-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
+// CHECK13:       omp_offload.cont30:
 // CHECK13-NEXT:    ret i32 0
 //
 //
@@ -7157,31 +7667,31 @@ int main (int argc, char **argv) {
 // CHECK14-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK14-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
 // CHECK14-NEXT:    [[M_CASTED:%.*]] = alloca i64, align 8
-// CHECK14-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [4 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [4 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [4 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_SIZES24:%.*]] = alloca [4 x i64], align 8
-// CHECK14-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
-// CHECK14-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK14-NEXT:    [[N_CASTED20:%.*]] = alloca i64, align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS22:%.*]] = alloca [4 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_PTRS23:%.*]] = alloca [4 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS24:%.*]] = alloca [4 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_SIZES25:%.*]] = alloca [4 x i64], align 8
+// CHECK14-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
 // CHECK14-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
-// CHECK14-NEXT:    [[N_CASTED34:%.*]] = alloca i64, align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS36:%.*]] = alloca [3 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_PTRS37:%.*]] = alloca [3 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS38:%.*]] = alloca [3 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_SIZES39:%.*]] = alloca [3 x i64], align 8
-// CHECK14-NEXT:    [[_TMP40:%.*]] = alloca i32, align 4
-// CHECK14-NEXT:    [[DOTCAPTURE_EXPR_41:%.*]] = alloca i32, align 4
-// CHECK14-NEXT:    [[DOTCAPTURE_EXPR_42:%.*]] = alloca i32, align 4
-// CHECK14-NEXT:    [[M_CASTED49:%.*]] = alloca i64, align 8
-// CHECK14-NEXT:    [[N_CASTED51:%.*]] = alloca i64, align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS53:%.*]] = alloca [4 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_PTRS54:%.*]] = alloca [4 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS55:%.*]] = alloca [4 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_SIZES56:%.*]] = alloca [4 x i64], align 8
-// CHECK14-NEXT:    [[_TMP57:%.*]] = alloca i32, align 4
-// CHECK14-NEXT:    [[DOTCAPTURE_EXPR_58:%.*]] = alloca i32, align 4
-// CHECK14-NEXT:    [[DOTCAPTURE_EXPR_59:%.*]] = alloca i32, align 4
+// CHECK14-NEXT:    [[DOTCAPTURE_EXPR_28:%.*]] = alloca i32, align 4
+// CHECK14-NEXT:    [[N_CASTED36:%.*]] = alloca i64, align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS38:%.*]] = alloca [3 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_PTRS39:%.*]] = alloca [3 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS40:%.*]] = alloca [3 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_SIZES41:%.*]] = alloca [3 x i64], align 8
+// CHECK14-NEXT:    [[_TMP42:%.*]] = alloca i32, align 4
+// CHECK14-NEXT:    [[DOTCAPTURE_EXPR_43:%.*]] = alloca i32, align 4
+// CHECK14-NEXT:    [[DOTCAPTURE_EXPR_44:%.*]] = alloca i32, align 4
+// CHECK14-NEXT:    [[M_CASTED52:%.*]] = alloca i64, align 8
+// CHECK14-NEXT:    [[N_CASTED54:%.*]] = alloca i64, align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS56:%.*]] = alloca [4 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_PTRS57:%.*]] = alloca [4 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS58:%.*]] = alloca [4 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_SIZES59:%.*]] = alloca [4 x i64], align 8
+// CHECK14-NEXT:    [[_TMP60:%.*]] = alloca i32, align 4
+// CHECK14-NEXT:    [[DOTCAPTURE_EXPR_61:%.*]] = alloca i32, align 4
+// CHECK14-NEXT:    [[DOTCAPTURE_EXPR_62:%.*]] = alloca i32, align 4
 // CHECK14-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK14-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK14-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -7240,260 +7750,345 @@ int main (int argc, char **argv) {
 // CHECK14-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK14-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK14-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK14-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK14-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK14-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK14-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK14-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK14-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK14-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK14-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK14-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK14-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK14-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK14-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK14-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK14-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK14-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK14-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK14-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK14-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK14-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK14:       omp_offload.failed:
 // CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK14:       omp_offload.cont:
-// CHECK14-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK14-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK14-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK14-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK14-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK14-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK14-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK14-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
-// CHECK14-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK14-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK14-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK14-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK14-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK14-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK14-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK14-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK14-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK14-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK14-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK14-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK14-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK14-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK14-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK14-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK14-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK14-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK14-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK14-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK14-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK14-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK14-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK14-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK14-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK14-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK14-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i64 24, i1 false)
+// CHECK14-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK14-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK14-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK14-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK14-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK14-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK14-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK14-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK14-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK14-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK14-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK14-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK14-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK14-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK14-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK14-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK14-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK14-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK14-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK14-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK14-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK14-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK14-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK14-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK14-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK14-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK14-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK14-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP58]])
-// CHECK14-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK14-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK14-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK14:       omp_offload.failed16:
-// CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK14:       omp_offload.cont17:
-// CHECK14-NEXT:    [[TMP61:%.*]] = load i32, i32* [[M]], align 4
-// CHECK14-NEXT:    [[CONV18:%.*]] = bitcast i64* [[M_CASTED]] to i32*
-// CHECK14-NEXT:    store i32 [[TMP61]], i32* [[CONV18]], align 4
-// CHECK14-NEXT:    [[TMP62:%.*]] = load i64, i64* [[M_CASTED]], align 8
-// CHECK14-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
-// CHECK14-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
-// CHECK14-NEXT:    store i32 [[TMP63]], i32* [[CONV20]], align 4
-// CHECK14-NEXT:    [[TMP64:%.*]] = load i64, i64* [[N_CASTED19]], align 8
-// CHECK14-NEXT:    [[TMP65:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK14-NEXT:    [[TMP66:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES24]] to i8*
-// CHECK14-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP66]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
-// CHECK14-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP62]], i64* [[TMP68]], align 8
-// CHECK14-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP62]], i64* [[TMP70]], align 8
-// CHECK14-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
-// CHECK14-NEXT:    store i8* null, i8** [[TMP71]], align 8
-// CHECK14-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK14-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP64]], i64* [[TMP73]], align 8
-// CHECK14-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK14-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP64]], i64* [[TMP75]], align 8
-// CHECK14-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
-// CHECK14-NEXT:    store i8* null, i8** [[TMP76]], align 8
-// CHECK14-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
-// CHECK14-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP78]], align 8
-// CHECK14-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
-// CHECK14-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP80]], align 8
-// CHECK14-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
-// CHECK14-NEXT:    store i8* null, i8** [[TMP81]], align 8
-// CHECK14-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 3
-// CHECK14-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to i32**
-// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP83]], align 8
-// CHECK14-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 3
-// CHECK14-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32**
-// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP85]], align 8
-// CHECK14-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 3
-// CHECK14-NEXT:    store i64 [[TMP65]], i64* [[TMP86]], align 8
-// CHECK14-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 3
+// CHECK14-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK14-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK14-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP66]])
+// CHECK14-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK14-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK14-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK14-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK14-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK14-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK14-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK14-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK14-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK14-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK14-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK14-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK14-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK14-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK14-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK14-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK14-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK14:       omp_offload.failed17:
+// CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK14:       omp_offload.cont18:
+// CHECK14-NEXT:    [[TMP77:%.*]] = load i32, i32* [[M]], align 4
+// CHECK14-NEXT:    [[CONV19:%.*]] = bitcast i64* [[M_CASTED]] to i32*
+// CHECK14-NEXT:    store i32 [[TMP77]], i32* [[CONV19]], align 4
+// CHECK14-NEXT:    [[TMP78:%.*]] = load i64, i64* [[M_CASTED]], align 8
+// CHECK14-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N]], align 4
+// CHECK14-NEXT:    [[CONV21:%.*]] = bitcast i64* [[N_CASTED20]] to i32*
+// CHECK14-NEXT:    store i32 [[TMP79]], i32* [[CONV21]], align 4
+// CHECK14-NEXT:    [[TMP80:%.*]] = load i64, i64* [[N_CASTED20]], align 8
+// CHECK14-NEXT:    [[TMP81:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK14-NEXT:    [[TMP82:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES25]] to i8*
+// CHECK14-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP82]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i64 32, i1 false)
+// CHECK14-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP78]], i64* [[TMP84]], align 8
+// CHECK14-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP78]], i64* [[TMP86]], align 8
+// CHECK14-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 0
 // CHECK14-NEXT:    store i8* null, i8** [[TMP87]], align 8
-// CHECK14-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP91:%.*]] = load i32, i32* [[N]], align 4
-// CHECK14-NEXT:    store i32 [[TMP91]], i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK14-NEXT:    [[TMP92:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK14-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[TMP92]], 0
-// CHECK14-NEXT:    [[DIV29:%.*]] = sdiv i32 [[SUB28]], 1
-// CHECK14-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[DIV29]], 1
-// CHECK14-NEXT:    store i32 [[SUB30]], i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK14-NEXT:    [[TMP93:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
-// CHECK14-NEXT:    [[ADD31:%.*]] = add nsw i32 [[TMP93]], 1
-// CHECK14-NEXT:    [[TMP94:%.*]] = zext i32 [[ADD31]] to i64
-// CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP94]])
-// CHECK14-NEXT:    [[TMP95:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164.region_id, i32 4, i8** [[TMP88]], i8** [[TMP89]], i64* [[TMP90]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK14-NEXT:    [[TMP96:%.*]] = icmp ne i32 [[TMP95]], 0
-// CHECK14-NEXT:    br i1 [[TMP96]], label [[OMP_OFFLOAD_FAILED32:%.*]], label [[OMP_OFFLOAD_CONT33:%.*]]
-// CHECK14:       omp_offload.failed32:
-// CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164(i64 [[TMP62]], i64 [[TMP64]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT33]]
-// CHECK14:       omp_offload.cont33:
-// CHECK14-NEXT:    [[TMP97:%.*]] = load i32, i32* [[N]], align 4
-// CHECK14-NEXT:    [[CONV35:%.*]] = bitcast i64* [[N_CASTED34]] to i32*
-// CHECK14-NEXT:    store i32 [[TMP97]], i32* [[CONV35]], align 4
-// CHECK14-NEXT:    [[TMP98:%.*]] = load i64, i64* [[N_CASTED34]], align 8
-// CHECK14-NEXT:    [[TMP99:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK14-NEXT:    [[TMP100:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES39]] to i8*
-// CHECK14-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP100]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
-// CHECK14-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP102:%.*]] = bitcast i8** [[TMP101]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP98]], i64* [[TMP102]], align 8
-// CHECK14-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP104:%.*]] = bitcast i8** [[TMP103]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP98]], i64* [[TMP104]], align 8
-// CHECK14-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 0
-// CHECK14-NEXT:    store i8* null, i8** [[TMP105]], align 8
-// CHECK14-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 1
-// CHECK14-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP107]], align 8
-// CHECK14-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 1
-// CHECK14-NEXT:    [[TMP109:%.*]] = bitcast i8** [[TMP108]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP109]], align 8
-// CHECK14-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 1
-// CHECK14-NEXT:    store i8* null, i8** [[TMP110]], align 8
-// CHECK14-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 2
-// CHECK14-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32**
-// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP112]], align 8
-// CHECK14-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 2
-// CHECK14-NEXT:    [[TMP114:%.*]] = bitcast i8** [[TMP113]] to i32**
-// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP114]], align 8
-// CHECK14-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES39]], i32 0, i32 2
-// CHECK14-NEXT:    store i64 [[TMP99]], i64* [[TMP115]], align 8
-// CHECK14-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS38]], i64 0, i64 2
-// CHECK14-NEXT:    store i8* null, i8** [[TMP116]], align 8
-// CHECK14-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS36]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS37]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES39]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP120:%.*]] = load i32, i32* [[N]], align 4
-// CHECK14-NEXT:    store i32 [[TMP120]], i32* [[DOTCAPTURE_EXPR_41]], align 4
-// CHECK14-NEXT:    [[TMP121:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_41]], align 4
-// CHECK14-NEXT:    [[SUB43:%.*]] = sub nsw i32 [[TMP121]], 0
-// CHECK14-NEXT:    [[DIV44:%.*]] = sdiv i32 [[SUB43]], 1
-// CHECK14-NEXT:    [[SUB45:%.*]] = sub nsw i32 [[DIV44]], 1
-// CHECK14-NEXT:    store i32 [[SUB45]], i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK14-NEXT:    [[TMP122:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_42]], align 4
-// CHECK14-NEXT:    [[ADD46:%.*]] = add nsw i32 [[TMP122]], 1
-// CHECK14-NEXT:    [[TMP123:%.*]] = zext i32 [[ADD46]] to i64
-// CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP123]])
-// CHECK14-NEXT:    [[TMP124:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169.region_id, i32 3, i8** [[TMP117]], i8** [[TMP118]], i64* [[TMP119]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK14-NEXT:    [[TMP125:%.*]] = icmp ne i32 [[TMP124]], 0
-// CHECK14-NEXT:    br i1 [[TMP125]], label [[OMP_OFFLOAD_FAILED47:%.*]], label [[OMP_OFFLOAD_CONT48:%.*]]
-// CHECK14:       omp_offload.failed47:
-// CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169(i64 [[TMP98]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT48]]
-// CHECK14:       omp_offload.cont48:
-// CHECK14-NEXT:    [[TMP126:%.*]] = load i32, i32* [[M]], align 4
-// CHECK14-NEXT:    [[CONV50:%.*]] = bitcast i64* [[M_CASTED49]] to i32*
-// CHECK14-NEXT:    store i32 [[TMP126]], i32* [[CONV50]], align 4
-// CHECK14-NEXT:    [[TMP127:%.*]] = load i64, i64* [[M_CASTED49]], align 8
-// CHECK14-NEXT:    [[TMP128:%.*]] = load i32, i32* [[N]], align 4
-// CHECK14-NEXT:    [[CONV52:%.*]] = bitcast i64* [[N_CASTED51]] to i32*
-// CHECK14-NEXT:    store i32 [[TMP128]], i32* [[CONV52]], align 4
-// CHECK14-NEXT:    [[TMP129:%.*]] = load i64, i64* [[N_CASTED51]], align 8
-// CHECK14-NEXT:    [[TMP130:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK14-NEXT:    [[TMP131:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES56]] to i8*
-// CHECK14-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP131]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
-// CHECK14-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 1
+// CHECK14-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP80]], i64* [[TMP89]], align 8
+// CHECK14-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 1
+// CHECK14-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP80]], i64* [[TMP91]], align 8
+// CHECK14-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 1
+// CHECK14-NEXT:    store i8* null, i8** [[TMP92]], align 8
+// CHECK14-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 2
+// CHECK14-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP94]], align 8
+// CHECK14-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 2
+// CHECK14-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP96]], align 8
+// CHECK14-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 2
+// CHECK14-NEXT:    store i8* null, i8** [[TMP97]], align 8
+// CHECK14-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 3
+// CHECK14-NEXT:    [[TMP99:%.*]] = bitcast i8** [[TMP98]] to i32**
+// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP99]], align 8
+// CHECK14-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 3
+// CHECK14-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32**
+// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP101]], align 8
+// CHECK14-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 3
+// CHECK14-NEXT:    store i64 [[TMP81]], i64* [[TMP102]], align 8
+// CHECK14-NEXT:    [[TMP103:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS24]], i64 0, i64 3
+// CHECK14-NEXT:    store i8* null, i8** [[TMP103]], align 8
+// CHECK14-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS22]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS23]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES25]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP107:%.*]] = load i32, i32* [[N]], align 4
+// CHECK14-NEXT:    store i32 [[TMP107]], i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK14-NEXT:    [[TMP108:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK14-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[TMP108]], 0
+// CHECK14-NEXT:    [[DIV30:%.*]] = sdiv i32 [[SUB29]], 1
+// CHECK14-NEXT:    [[SUB31:%.*]] = sub nsw i32 [[DIV30]], 1
+// CHECK14-NEXT:    store i32 [[SUB31]], i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK14-NEXT:    [[TMP109:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_28]], align 4
+// CHECK14-NEXT:    [[ADD32:%.*]] = add nsw i32 [[TMP109]], 1
+// CHECK14-NEXT:    [[TMP110:%.*]] = zext i32 [[ADD32]] to i64
+// CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP110]])
+// CHECK14-NEXT:    [[KERNEL_ARGS33:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK14-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 0
+// CHECK14-NEXT:    store i32 1, i32* [[TMP111]], align 4
+// CHECK14-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 1
+// CHECK14-NEXT:    store i32 4, i32* [[TMP112]], align 4
+// CHECK14-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 2
+// CHECK14-NEXT:    store i8** [[TMP104]], i8*** [[TMP113]], align 8
+// CHECK14-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 3
+// CHECK14-NEXT:    store i8** [[TMP105]], i8*** [[TMP114]], align 8
+// CHECK14-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 4
+// CHECK14-NEXT:    store i64* [[TMP106]], i64** [[TMP115]], align 8
+// CHECK14-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 5
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP116]], align 8
+// CHECK14-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 6
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP117]], align 8
+// CHECK14-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]], i32 0, i32 7
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP118]], align 8
+// CHECK14-NEXT:    [[TMP119:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS33]])
+// CHECK14-NEXT:    [[TMP120:%.*]] = icmp ne i32 [[TMP119]], 0
+// CHECK14-NEXT:    br i1 [[TMP120]], label [[OMP_OFFLOAD_FAILED34:%.*]], label [[OMP_OFFLOAD_CONT35:%.*]]
+// CHECK14:       omp_offload.failed34:
+// CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164(i64 [[TMP78]], i64 [[TMP80]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT35]]
+// CHECK14:       omp_offload.cont35:
+// CHECK14-NEXT:    [[TMP121:%.*]] = load i32, i32* [[N]], align 4
+// CHECK14-NEXT:    [[CONV37:%.*]] = bitcast i64* [[N_CASTED36]] to i32*
+// CHECK14-NEXT:    store i32 [[TMP121]], i32* [[CONV37]], align 4
+// CHECK14-NEXT:    [[TMP122:%.*]] = load i64, i64* [[N_CASTED36]], align 8
+// CHECK14-NEXT:    [[TMP123:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK14-NEXT:    [[TMP124:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES41]] to i8*
+// CHECK14-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP124]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i64 24, i1 false)
+// CHECK14-NEXT:    [[TMP125:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP126:%.*]] = bitcast i8** [[TMP125]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP122]], i64* [[TMP126]], align 8
+// CHECK14-NEXT:    [[TMP127:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP128:%.*]] = bitcast i8** [[TMP127]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP122]], i64* [[TMP128]], align 8
+// CHECK14-NEXT:    [[TMP129:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS40]], i64 0, i64 0
+// CHECK14-NEXT:    store i8* null, i8** [[TMP129]], align 8
+// CHECK14-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 1
+// CHECK14-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP131]], align 8
+// CHECK14-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 1
 // CHECK14-NEXT:    [[TMP133:%.*]] = bitcast i8** [[TMP132]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP127]], i64* [[TMP133]], align 8
-// CHECK14-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP135:%.*]] = bitcast i8** [[TMP134]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP127]], i64* [[TMP135]], align 8
-// CHECK14-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 0
-// CHECK14-NEXT:    store i8* null, i8** [[TMP136]], align 8
-// CHECK14-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 1
-// CHECK14-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP129]], i64* [[TMP138]], align 8
-// CHECK14-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 1
-// CHECK14-NEXT:    [[TMP140:%.*]] = bitcast i8** [[TMP139]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP129]], i64* [[TMP140]], align 8
-// CHECK14-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 1
-// CHECK14-NEXT:    store i8* null, i8** [[TMP141]], align 8
-// CHECK14-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 2
-// CHECK14-NEXT:    [[TMP143:%.*]] = bitcast i8** [[TMP142]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP143]], align 8
-// CHECK14-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 2
-// CHECK14-NEXT:    [[TMP145:%.*]] = bitcast i8** [[TMP144]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP145]], align 8
-// CHECK14-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 2
-// CHECK14-NEXT:    store i8* null, i8** [[TMP146]], align 8
-// CHECK14-NEXT:    [[TMP147:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 3
-// CHECK14-NEXT:    [[TMP148:%.*]] = bitcast i8** [[TMP147]] to i32**
-// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP148]], align 8
-// CHECK14-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 3
-// CHECK14-NEXT:    [[TMP150:%.*]] = bitcast i8** [[TMP149]] to i32**
-// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP150]], align 8
-// CHECK14-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES56]], i32 0, i32 3
-// CHECK14-NEXT:    store i64 [[TMP130]], i64* [[TMP151]], align 8
-// CHECK14-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS55]], i64 0, i64 3
-// CHECK14-NEXT:    store i8* null, i8** [[TMP152]], align 8
-// CHECK14-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS53]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS54]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES56]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP156:%.*]] = load i32, i32* [[N]], align 4
-// CHECK14-NEXT:    store i32 [[TMP156]], i32* [[DOTCAPTURE_EXPR_58]], align 4
-// CHECK14-NEXT:    [[TMP157:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_58]], align 4
-// CHECK14-NEXT:    [[SUB60:%.*]] = sub nsw i32 [[TMP157]], 0
-// CHECK14-NEXT:    [[DIV61:%.*]] = sdiv i32 [[SUB60]], 1
-// CHECK14-NEXT:    [[SUB62:%.*]] = sub nsw i32 [[DIV61]], 1
-// CHECK14-NEXT:    store i32 [[SUB62]], i32* [[DOTCAPTURE_EXPR_59]], align 4
-// CHECK14-NEXT:    [[TMP158:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_59]], align 4
-// CHECK14-NEXT:    [[ADD63:%.*]] = add nsw i32 [[TMP158]], 1
-// CHECK14-NEXT:    [[TMP159:%.*]] = zext i32 [[ADD63]] to i64
-// CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP159]])
-// CHECK14-NEXT:    [[TMP160:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174.region_id, i32 4, i8** [[TMP153]], i8** [[TMP154]], i64* [[TMP155]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK14-NEXT:    [[TMP161:%.*]] = icmp ne i32 [[TMP160]], 0
-// CHECK14-NEXT:    br i1 [[TMP161]], label [[OMP_OFFLOAD_FAILED64:%.*]], label [[OMP_OFFLOAD_CONT65:%.*]]
-// CHECK14:       omp_offload.failed64:
-// CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174(i64 [[TMP127]], i64 [[TMP129]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT65]]
-// CHECK14:       omp_offload.cont65:
-// CHECK14-NEXT:    [[TMP162:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK14-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP162]])
+// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP133]], align 8
+// CHECK14-NEXT:    [[TMP134:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS40]], i64 0, i64 1
+// CHECK14-NEXT:    store i8* null, i8** [[TMP134]], align 8
+// CHECK14-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 2
+// CHECK14-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32**
+// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP136]], align 8
+// CHECK14-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 2
+// CHECK14-NEXT:    [[TMP138:%.*]] = bitcast i8** [[TMP137]] to i32**
+// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP138]], align 8
+// CHECK14-NEXT:    [[TMP139:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES41]], i32 0, i32 2
+// CHECK14-NEXT:    store i64 [[TMP123]], i64* [[TMP139]], align 8
+// CHECK14-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS40]], i64 0, i64 2
+// CHECK14-NEXT:    store i8* null, i8** [[TMP140]], align 8
+// CHECK14-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS38]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS39]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES41]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP144:%.*]] = load i32, i32* [[N]], align 4
+// CHECK14-NEXT:    store i32 [[TMP144]], i32* [[DOTCAPTURE_EXPR_43]], align 4
+// CHECK14-NEXT:    [[TMP145:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_43]], align 4
+// CHECK14-NEXT:    [[SUB45:%.*]] = sub nsw i32 [[TMP145]], 0
+// CHECK14-NEXT:    [[DIV46:%.*]] = sdiv i32 [[SUB45]], 1
+// CHECK14-NEXT:    [[SUB47:%.*]] = sub nsw i32 [[DIV46]], 1
+// CHECK14-NEXT:    store i32 [[SUB47]], i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK14-NEXT:    [[TMP146:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_44]], align 4
+// CHECK14-NEXT:    [[ADD48:%.*]] = add nsw i32 [[TMP146]], 1
+// CHECK14-NEXT:    [[TMP147:%.*]] = zext i32 [[ADD48]] to i64
+// CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP147]])
+// CHECK14-NEXT:    [[KERNEL_ARGS49:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK14-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 0
+// CHECK14-NEXT:    store i32 1, i32* [[TMP148]], align 4
+// CHECK14-NEXT:    [[TMP149:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 1
+// CHECK14-NEXT:    store i32 3, i32* [[TMP149]], align 4
+// CHECK14-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 2
+// CHECK14-NEXT:    store i8** [[TMP141]], i8*** [[TMP150]], align 8
+// CHECK14-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 3
+// CHECK14-NEXT:    store i8** [[TMP142]], i8*** [[TMP151]], align 8
+// CHECK14-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 4
+// CHECK14-NEXT:    store i64* [[TMP143]], i64** [[TMP152]], align 8
+// CHECK14-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 5
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP153]], align 8
+// CHECK14-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 6
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP154]], align 8
+// CHECK14-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]], i32 0, i32 7
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP155]], align 8
+// CHECK14-NEXT:    [[TMP156:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS49]])
+// CHECK14-NEXT:    [[TMP157:%.*]] = icmp ne i32 [[TMP156]], 0
+// CHECK14-NEXT:    br i1 [[TMP157]], label [[OMP_OFFLOAD_FAILED50:%.*]], label [[OMP_OFFLOAD_CONT51:%.*]]
+// CHECK14:       omp_offload.failed50:
+// CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169(i64 [[TMP122]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT51]]
+// CHECK14:       omp_offload.cont51:
+// CHECK14-NEXT:    [[TMP158:%.*]] = load i32, i32* [[M]], align 4
+// CHECK14-NEXT:    [[CONV53:%.*]] = bitcast i64* [[M_CASTED52]] to i32*
+// CHECK14-NEXT:    store i32 [[TMP158]], i32* [[CONV53]], align 4
+// CHECK14-NEXT:    [[TMP159:%.*]] = load i64, i64* [[M_CASTED52]], align 8
+// CHECK14-NEXT:    [[TMP160:%.*]] = load i32, i32* [[N]], align 4
+// CHECK14-NEXT:    [[CONV55:%.*]] = bitcast i64* [[N_CASTED54]] to i32*
+// CHECK14-NEXT:    store i32 [[TMP160]], i32* [[CONV55]], align 4
+// CHECK14-NEXT:    [[TMP161:%.*]] = load i64, i64* [[N_CASTED54]], align 8
+// CHECK14-NEXT:    [[TMP162:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK14-NEXT:    [[TMP163:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES59]] to i8*
+// CHECK14-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP163]], i8* align 8 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i64 32, i1 false)
+// CHECK14-NEXT:    [[TMP164:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP165:%.*]] = bitcast i8** [[TMP164]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP159]], i64* [[TMP165]], align 8
+// CHECK14-NEXT:    [[TMP166:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP167:%.*]] = bitcast i8** [[TMP166]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP159]], i64* [[TMP167]], align 8
+// CHECK14-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 0
+// CHECK14-NEXT:    store i8* null, i8** [[TMP168]], align 8
+// CHECK14-NEXT:    [[TMP169:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 1
+// CHECK14-NEXT:    [[TMP170:%.*]] = bitcast i8** [[TMP169]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP161]], i64* [[TMP170]], align 8
+// CHECK14-NEXT:    [[TMP171:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 1
+// CHECK14-NEXT:    [[TMP172:%.*]] = bitcast i8** [[TMP171]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP161]], i64* [[TMP172]], align 8
+// CHECK14-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 1
+// CHECK14-NEXT:    store i8* null, i8** [[TMP173]], align 8
+// CHECK14-NEXT:    [[TMP174:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 2
+// CHECK14-NEXT:    [[TMP175:%.*]] = bitcast i8** [[TMP174]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP175]], align 8
+// CHECK14-NEXT:    [[TMP176:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 2
+// CHECK14-NEXT:    [[TMP177:%.*]] = bitcast i8** [[TMP176]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP1]], i64* [[TMP177]], align 8
+// CHECK14-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 2
+// CHECK14-NEXT:    store i8* null, i8** [[TMP178]], align 8
+// CHECK14-NEXT:    [[TMP179:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 3
+// CHECK14-NEXT:    [[TMP180:%.*]] = bitcast i8** [[TMP179]] to i32**
+// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP180]], align 8
+// CHECK14-NEXT:    [[TMP181:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 3
+// CHECK14-NEXT:    [[TMP182:%.*]] = bitcast i8** [[TMP181]] to i32**
+// CHECK14-NEXT:    store i32* [[VLA]], i32** [[TMP182]], align 8
+// CHECK14-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES59]], i32 0, i32 3
+// CHECK14-NEXT:    store i64 [[TMP162]], i64* [[TMP183]], align 8
+// CHECK14-NEXT:    [[TMP184:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS58]], i64 0, i64 3
+// CHECK14-NEXT:    store i8* null, i8** [[TMP184]], align 8
+// CHECK14-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS56]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP186:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS57]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES59]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP188:%.*]] = load i32, i32* [[N]], align 4
+// CHECK14-NEXT:    store i32 [[TMP188]], i32* [[DOTCAPTURE_EXPR_61]], align 4
+// CHECK14-NEXT:    [[TMP189:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_61]], align 4
+// CHECK14-NEXT:    [[SUB63:%.*]] = sub nsw i32 [[TMP189]], 0
+// CHECK14-NEXT:    [[DIV64:%.*]] = sdiv i32 [[SUB63]], 1
+// CHECK14-NEXT:    [[SUB65:%.*]] = sub nsw i32 [[DIV64]], 1
+// CHECK14-NEXT:    store i32 [[SUB65]], i32* [[DOTCAPTURE_EXPR_62]], align 4
+// CHECK14-NEXT:    [[TMP190:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_62]], align 4
+// CHECK14-NEXT:    [[ADD66:%.*]] = add nsw i32 [[TMP190]], 1
+// CHECK14-NEXT:    [[TMP191:%.*]] = zext i32 [[ADD66]] to i64
+// CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP191]])
+// CHECK14-NEXT:    [[KERNEL_ARGS67:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK14-NEXT:    [[TMP192:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 0
+// CHECK14-NEXT:    store i32 1, i32* [[TMP192]], align 4
+// CHECK14-NEXT:    [[TMP193:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 1
+// CHECK14-NEXT:    store i32 4, i32* [[TMP193]], align 4
+// CHECK14-NEXT:    [[TMP194:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 2
+// CHECK14-NEXT:    store i8** [[TMP185]], i8*** [[TMP194]], align 8
+// CHECK14-NEXT:    [[TMP195:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 3
+// CHECK14-NEXT:    store i8** [[TMP186]], i8*** [[TMP195]], align 8
+// CHECK14-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 4
+// CHECK14-NEXT:    store i64* [[TMP187]], i64** [[TMP196]], align 8
+// CHECK14-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 5
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP197]], align 8
+// CHECK14-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 6
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP198]], align 8
+// CHECK14-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]], i32 0, i32 7
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP199]], align 8
+// CHECK14-NEXT:    [[TMP200:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS67]])
+// CHECK14-NEXT:    [[TMP201:%.*]] = icmp ne i32 [[TMP200]], 0
+// CHECK14-NEXT:    br i1 [[TMP201]], label [[OMP_OFFLOAD_FAILED68:%.*]], label [[OMP_OFFLOAD_CONT69:%.*]]
+// CHECK14:       omp_offload.failed68:
+// CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174(i64 [[TMP159]], i64 [[TMP161]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT69]]
+// CHECK14:       omp_offload.cont69:
+// CHECK14-NEXT:    [[TMP202:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK14-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP202]])
 // CHECK14-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK14-NEXT:    [[TMP163:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK14-NEXT:    call void @llvm.stackrestore(i8* [[TMP163]])
-// CHECK14-NEXT:    [[TMP164:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK14-NEXT:    ret i32 [[TMP164]]
+// CHECK14-NEXT:    [[TMP203:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK14-NEXT:    call void @llvm.stackrestore(i8* [[TMP203]])
+// CHECK14-NEXT:    [[TMP204:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK14-NEXT:    ret i32 [[TMP204]]
 //
 //
 // CHECK14-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154
@@ -8753,19 +9348,19 @@ int main (int argc, char **argv) {
 // CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 8
 // CHECK14-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK14-NEXT:    [[M_CASTED:%.*]] = alloca i64, align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 8
-// CHECK14-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 8
-// CHECK14-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK14-NEXT:    [[M_CASTED19:%.*]] = alloca i64, align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [2 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [2 x i8*], align 8
-// CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [2 x i8*], align 8
-// CHECK14-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
+// CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 8
+// CHECK14-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 8
+// CHECK14-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK14-NEXT:    [[M_CASTED22:%.*]] = alloca i64, align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_BASEPTRS24:%.*]] = alloca [2 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_PTRS25:%.*]] = alloca [2 x i8*], align 8
+// CHECK14-NEXT:    [[DOTOFFLOAD_MAPPERS26:%.*]] = alloca [2 x i8*], align 8
+// CHECK14-NEXT:    [[_TMP27:%.*]] = alloca i32, align 4
 // CHECK14-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK14-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK14-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -8779,109 +9374,194 @@ int main (int argc, char **argv) {
 // CHECK14-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK14-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK14-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l122.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK14-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK14-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK14-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK14-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK14-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK14-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK14-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK14-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK14-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK14-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK14-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK14-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK14-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK14-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK14-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK14-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l122.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK14-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK14-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK14:       omp_offload.failed:
 // CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l122([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK14:       omp_offload.cont:
-// CHECK14-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK14-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK14-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK14-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK14-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK14-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK14-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK14-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK14-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK14-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l127.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK14-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK14-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK14:       omp_offload.failed5:
+// CHECK14-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK14-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK14-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK14-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK14-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK14-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK14-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK14-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK14-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK14-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK14-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK14-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK14-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK14-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l127.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK14-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK14-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK14:       omp_offload.failed6:
 // CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l127([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK14:       omp_offload.cont6:
-// CHECK14-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
+// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK14:       omp_offload.cont7:
+// CHECK14-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
 // CHECK14-NEXT:    [[CONV:%.*]] = bitcast i64* [[M_CASTED]] to i32*
-// CHECK14-NEXT:    store i32 [[TMP18]], i32* [[CONV]], align 4
-// CHECK14-NEXT:    [[TMP19:%.*]] = load i64, i64* [[M_CASTED]], align 8
-// CHECK14-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP19]], i64* [[TMP21]], align 8
-// CHECK14-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP19]], i64* [[TMP23]], align 8
-// CHECK14-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK14-NEXT:    store i8* null, i8** [[TMP24]], align 8
-// CHECK14-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK14-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to [10 x i32]**
-// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP26]], align 8
-// CHECK14-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK14-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [10 x i32]**
-// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP28]], align 8
-// CHECK14-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 1
-// CHECK14-NEXT:    store i8* null, i8** [[TMP29]], align 8
-// CHECK14-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK14-NEXT:    store i32 [[TMP34]], i32* [[CONV]], align 4
+// CHECK14-NEXT:    [[TMP35:%.*]] = load i64, i64* [[M_CASTED]], align 8
+// CHECK14-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP35]], i64* [[TMP37]], align 8
+// CHECK14-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP35]], i64* [[TMP39]], align 8
+// CHECK14-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK14-NEXT:    store i8* null, i8** [[TMP40]], align 8
+// CHECK14-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK14-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to [10 x i32]**
+// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP42]], align 8
+// CHECK14-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK14-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [10 x i32]**
+// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP44]], align 8
+// CHECK14-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 1
+// CHECK14-NEXT:    store i8* null, i8** [[TMP45]], align 8
+// CHECK14-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK14-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132.region_id, i32 2, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK14-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK14-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK14:       omp_offload.failed11:
-// CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132(i64 [[TMP19]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK14:       omp_offload.cont12:
-// CHECK14-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
-// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 8
-// CHECK14-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
-// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 8
-// CHECK14-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i64 0, i64 0
-// CHECK14-NEXT:    store i8* null, i8** [[TMP38]], align 8
-// CHECK14-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK14-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK14-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK14-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK14-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK14-NEXT:    store i32 2, i32* [[TMP49]], align 4
+// CHECK14-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK14-NEXT:    store i8** [[TMP46]], i8*** [[TMP50]], align 8
+// CHECK14-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK14-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 8
+// CHECK14-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK14-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP53]], align 8
+// CHECK14-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK14-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP55]], align 8
+// CHECK14-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK14-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK14-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK14:       omp_offload.failed13:
+// CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132(i64 [[TMP35]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK14:       omp_offload.cont14:
+// CHECK14-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to [10 x i32]**
+// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP59]], align 8
+// CHECK14-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to [10 x i32]**
+// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP61]], align 8
+// CHECK14-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i64 0, i64 0
+// CHECK14-NEXT:    store i8* null, i8** [[TMP62]], align 8
+// CHECK14-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK14-NEXT:    [[TMP41:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l137.region_id, i32 1, i8** [[TMP39]], i8** [[TMP40]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK14-NEXT:    [[TMP42:%.*]] = icmp ne i32 [[TMP41]], 0
-// CHECK14-NEXT:    br i1 [[TMP42]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK14:       omp_offload.failed17:
+// CHECK14-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK14-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK14-NEXT:    store i32 1, i32* [[TMP65]], align 4
+// CHECK14-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK14-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK14-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK14-NEXT:    store i8** [[TMP63]], i8*** [[TMP67]], align 8
+// CHECK14-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK14-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 8
+// CHECK14-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP69]], align 8
+// CHECK14-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP70]], align 8
+// CHECK14-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP71]], align 8
+// CHECK14-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP72]], align 8
+// CHECK14-NEXT:    [[TMP73:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l137.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK14-NEXT:    [[TMP74:%.*]] = icmp ne i32 [[TMP73]], 0
+// CHECK14-NEXT:    br i1 [[TMP74]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK14:       omp_offload.failed20:
 // CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l137([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK14:       omp_offload.cont18:
-// CHECK14-NEXT:    [[TMP43:%.*]] = load i32, i32* [[M]], align 4
-// CHECK14-NEXT:    [[CONV20:%.*]] = bitcast i64* [[M_CASTED19]] to i32*
-// CHECK14-NEXT:    store i32 [[TMP43]], i32* [[CONV20]], align 4
-// CHECK14-NEXT:    [[TMP44:%.*]] = load i64, i64* [[M_CASTED19]], align 8
-// CHECK14-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP44]], i64* [[TMP46]], align 8
-// CHECK14-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i64*
-// CHECK14-NEXT:    store i64 [[TMP44]], i64* [[TMP48]], align 8
-// CHECK14-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
-// CHECK14-NEXT:    store i8* null, i8** [[TMP49]], align 8
-// CHECK14-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
-// CHECK14-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to [10 x i32]**
-// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP51]], align 8
-// CHECK14-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
-// CHECK14-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to [10 x i32]**
-// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP53]], align 8
-// CHECK14-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
-// CHECK14-NEXT:    store i8* null, i8** [[TMP54]], align 8
-// CHECK14-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
-// CHECK14-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK14:       omp_offload.cont21:
+// CHECK14-NEXT:    [[TMP75:%.*]] = load i32, i32* [[M]], align 4
+// CHECK14-NEXT:    [[CONV23:%.*]] = bitcast i64* [[M_CASTED22]] to i32*
+// CHECK14-NEXT:    store i32 [[TMP75]], i32* [[CONV23]], align 4
+// CHECK14-NEXT:    [[TMP76:%.*]] = load i64, i64* [[M_CASTED22]], align 8
+// CHECK14-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP76]], i64* [[TMP78]], align 8
+// CHECK14-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i64*
+// CHECK14-NEXT:    store i64 [[TMP76]], i64* [[TMP80]], align 8
+// CHECK14-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 0
+// CHECK14-NEXT:    store i8* null, i8** [[TMP81]], align 8
+// CHECK14-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 1
+// CHECK14-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to [10 x i32]**
+// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP83]], align 8
+// CHECK14-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 1
+// CHECK14-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to [10 x i32]**
+// CHECK14-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP85]], align 8
+// CHECK14-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS26]], i64 0, i64 1
+// CHECK14-NEXT:    store i8* null, i8** [[TMP86]], align 8
+// CHECK14-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS24]], i32 0, i32 0
+// CHECK14-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS25]], i32 0, i32 0
 // CHECK14-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK14-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK14-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
-// CHECK14-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED25:%.*]], label [[OMP_OFFLOAD_CONT26:%.*]]
-// CHECK14:       omp_offload.failed25:
-// CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142(i64 [[TMP44]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT26]]
-// CHECK14:       omp_offload.cont26:
+// CHECK14-NEXT:    [[KERNEL_ARGS28:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK14-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 0
+// CHECK14-NEXT:    store i32 1, i32* [[TMP89]], align 4
+// CHECK14-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 1
+// CHECK14-NEXT:    store i32 2, i32* [[TMP90]], align 4
+// CHECK14-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 2
+// CHECK14-NEXT:    store i8** [[TMP87]], i8*** [[TMP91]], align 8
+// CHECK14-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 3
+// CHECK14-NEXT:    store i8** [[TMP88]], i8*** [[TMP92]], align 8
+// CHECK14-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 4
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP93]], align 8
+// CHECK14-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 5
+// CHECK14-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP94]], align 8
+// CHECK14-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 6
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP95]], align 8
+// CHECK14-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]], i32 0, i32 7
+// CHECK14-NEXT:    store i8** null, i8*** [[TMP96]], align 8
+// CHECK14-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS28]])
+// CHECK14-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
+// CHECK14-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
+// CHECK14:       omp_offload.failed29:
+// CHECK14-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142(i64 [[TMP76]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK14-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
+// CHECK14:       omp_offload.cont30:
 // CHECK14-NEXT:    ret i32 0
 //
 //
@@ -9802,31 +10482,31 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[M_CASTED:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[N_CASTED17:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS18:%.*]] = alloca [4 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES21:%.*]] = alloca [4 x i64], align 4
-// CHECK17-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_23:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
+// CHECK17-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[N_CASTED31:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS32:%.*]] = alloca [3 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS33:%.*]] = alloca [3 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS34:%.*]] = alloca [3 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES35:%.*]] = alloca [3 x i64], align 4
-// CHECK17-NEXT:    [[_TMP36:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_37:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_38:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[M_CASTED45:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[N_CASTED46:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS47:%.*]] = alloca [4 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS48:%.*]] = alloca [4 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS49:%.*]] = alloca [4 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES50:%.*]] = alloca [4 x i64], align 4
-// CHECK17-NEXT:    [[_TMP51:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_52:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_53:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[N_CASTED33:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS34:%.*]] = alloca [3 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS35:%.*]] = alloca [3 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS36:%.*]] = alloca [3 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES37:%.*]] = alloca [3 x i64], align 4
+// CHECK17-NEXT:    [[_TMP38:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_39:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_40:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[M_CASTED48:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[N_CASTED49:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS50:%.*]] = alloca [4 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS51:%.*]] = alloca [4 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS52:%.*]] = alloca [4 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_SIZES53:%.*]] = alloca [4 x i64], align 4
+// CHECK17-NEXT:    [[_TMP54:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_55:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTCAPTURE_EXPR_56:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK17-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK17-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -9884,258 +10564,343 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK17-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK17-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK17-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK17-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK17-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK17-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK17-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK17-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK17-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK17-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
-// CHECK17-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK17-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK17-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK17-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK17-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
-// CHECK17-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK17-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK17-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK17-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK17-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK17-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK17-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK17-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK17-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK17-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK17-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK17-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK17-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK17-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK17-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
+// CHECK17-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK17-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK17-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK17-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK17-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK17-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK17-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK17-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK17-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK17-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK17-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK17-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK17-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK17-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK17-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK17-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK17-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP59]])
-// CHECK17-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK17-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK17:       omp_offload.failed15:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK17:       omp_offload.cont16:
-// CHECK17-NEXT:    [[TMP62:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    store i32 [[TMP62]], i32* [[M_CASTED]], align 4
-// CHECK17-NEXT:    [[TMP63:%.*]] = load i32, i32* [[M_CASTED]], align 4
-// CHECK17-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP64]], i32* [[N_CASTED17]], align 4
-// CHECK17-NEXT:    [[TMP65:%.*]] = load i32, i32* [[N_CASTED17]], align 4
-// CHECK17-NEXT:    [[TMP66:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK17-NEXT:    [[TMP67:%.*]] = sext i32 [[TMP66]] to i64
-// CHECK17-NEXT:    [[TMP68:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES21]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP68]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
-// CHECK17-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP63]], i32* [[TMP70]], align 4
-// CHECK17-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP63]], i32* [[TMP72]], align 4
-// CHECK17-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP73]], align 4
-// CHECK17-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP65]], i32* [[TMP75]], align 4
-// CHECK17-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP65]], i32* [[TMP77]], align 4
-// CHECK17-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP78]], align 4
-// CHECK17-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP80]], align 4
-// CHECK17-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP82]], align 4
-// CHECK17-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP83]], align 4
-// CHECK17-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP85]], align 4
-// CHECK17-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP87]], align 4
-// CHECK17-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 3
-// CHECK17-NEXT:    store i64 [[TMP67]], i64* [[TMP88]], align 4
-// CHECK17-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK17-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK17-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP67]])
+// CHECK17-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK17-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK17-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK17-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK17-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK17-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK17-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK17-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK17-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK17-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK17-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK17:       omp_offload.failed16:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK17:       omp_offload.cont17:
+// CHECK17-NEXT:    [[TMP78:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    store i32 [[TMP78]], i32* [[M_CASTED]], align 4
+// CHECK17-NEXT:    [[TMP79:%.*]] = load i32, i32* [[M_CASTED]], align 4
+// CHECK17-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP80]], i32* [[N_CASTED18]], align 4
+// CHECK17-NEXT:    [[TMP81:%.*]] = load i32, i32* [[N_CASTED18]], align 4
+// CHECK17-NEXT:    [[TMP82:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK17-NEXT:    [[TMP83:%.*]] = sext i32 [[TMP82]] to i64
+// CHECK17-NEXT:    [[TMP84:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP84]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
+// CHECK17-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP79]], i32* [[TMP86]], align 4
+// CHECK17-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP88:%.*]] = bitcast i8** [[TMP87]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP79]], i32* [[TMP88]], align 4
+// CHECK17-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
 // CHECK17-NEXT:    store i8* null, i8** [[TMP89]], align 4
-// CHECK17-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP93:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP93]], i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK17-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK17-NEXT:    [[SUB25:%.*]] = sub nsw i32 [[TMP94]], 0
-// CHECK17-NEXT:    [[DIV26:%.*]] = sdiv i32 [[SUB25]], 1
-// CHECK17-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[DIV26]], 1
-// CHECK17-NEXT:    store i32 [[SUB27]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK17-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK17-NEXT:    [[ADD28:%.*]] = add nsw i32 [[TMP95]], 1
-// CHECK17-NEXT:    [[TMP96:%.*]] = zext i32 [[ADD28]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP96]])
-// CHECK17-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164.region_id, i32 4, i8** [[TMP90]], i8** [[TMP91]], i64* [[TMP92]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
-// CHECK17-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
-// CHECK17:       omp_offload.failed29:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164(i32 [[TMP63]], i32 [[TMP65]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
-// CHECK17:       omp_offload.cont30:
-// CHECK17-NEXT:    [[TMP99:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP99]], i32* [[N_CASTED31]], align 4
-// CHECK17-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N_CASTED31]], align 4
-// CHECK17-NEXT:    [[TMP101:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK17-NEXT:    [[TMP102:%.*]] = sext i32 [[TMP101]] to i64
-// CHECK17-NEXT:    [[TMP103:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES35]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP103]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
-// CHECK17-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP100]], i32* [[TMP105]], align 4
-// CHECK17-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP100]], i32* [[TMP107]], align 4
-// CHECK17-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP108]], align 4
-// CHECK17-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP110]], align 4
-// CHECK17-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP112]], align 4
-// CHECK17-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP113]], align 4
-// CHECK17-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP115]], align 4
-// CHECK17-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP117]], align 4
-// CHECK17-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES35]], i32 0, i32 2
-// CHECK17-NEXT:    store i64 [[TMP102]], i64* [[TMP118]], align 4
-// CHECK17-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP119]], align 4
-// CHECK17-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES35]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP81]], i32* [[TMP91]], align 4
+// CHECK17-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP81]], i32* [[TMP93]], align 4
+// CHECK17-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP94]], align 4
+// CHECK17-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP96]], align 4
+// CHECK17-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP98]], align 4
+// CHECK17-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP99]], align 4
+// CHECK17-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP101]], align 4
+// CHECK17-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP103]], align 4
+// CHECK17-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 3
+// CHECK17-NEXT:    store i64 [[TMP83]], i64* [[TMP104]], align 4
+// CHECK17-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK17-NEXT:    store i8* null, i8** [[TMP105]], align 4
+// CHECK17-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP109:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP109]], i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK17-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK17-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP110]], 0
+// CHECK17-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
+// CHECK17-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
+// CHECK17-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK17-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK17-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP111]], 1
+// CHECK17-NEXT:    [[TMP112:%.*]] = zext i32 [[ADD29]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP112]])
+// CHECK17-NEXT:    [[KERNEL_ARGS30:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP113]], align 4
+// CHECK17-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 4, i32* [[TMP114]], align 4
+// CHECK17-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 4
+// CHECK17-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK17-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP108]], i64** [[TMP117]], align 4
+// CHECK17-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP118]], align 4
+// CHECK17-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP119]], align 4
+// CHECK17-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK17-NEXT:    [[TMP121:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]])
+// CHECK17-NEXT:    [[TMP122:%.*]] = icmp ne i32 [[TMP121]], 0
+// CHECK17-NEXT:    br i1 [[TMP122]], label [[OMP_OFFLOAD_FAILED31:%.*]], label [[OMP_OFFLOAD_CONT32:%.*]]
+// CHECK17:       omp_offload.failed31:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164(i32 [[TMP79]], i32 [[TMP81]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT32]]
+// CHECK17:       omp_offload.cont32:
 // CHECK17-NEXT:    [[TMP123:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP123]], i32* [[DOTCAPTURE_EXPR_37]], align 4
-// CHECK17-NEXT:    [[TMP124:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_37]], align 4
-// CHECK17-NEXT:    [[SUB39:%.*]] = sub nsw i32 [[TMP124]], 0
-// CHECK17-NEXT:    [[DIV40:%.*]] = sdiv i32 [[SUB39]], 1
-// CHECK17-NEXT:    [[SUB41:%.*]] = sub nsw i32 [[DIV40]], 1
-// CHECK17-NEXT:    store i32 [[SUB41]], i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK17-NEXT:    [[TMP125:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK17-NEXT:    [[ADD42:%.*]] = add nsw i32 [[TMP125]], 1
-// CHECK17-NEXT:    [[TMP126:%.*]] = zext i32 [[ADD42]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP126]])
-// CHECK17-NEXT:    [[TMP127:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169.region_id, i32 3, i8** [[TMP120]], i8** [[TMP121]], i64* [[TMP122]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP128:%.*]] = icmp ne i32 [[TMP127]], 0
-// CHECK17-NEXT:    br i1 [[TMP128]], label [[OMP_OFFLOAD_FAILED43:%.*]], label [[OMP_OFFLOAD_CONT44:%.*]]
-// CHECK17:       omp_offload.failed43:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169(i32 [[TMP100]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT44]]
-// CHECK17:       omp_offload.cont44:
-// CHECK17-NEXT:    [[TMP129:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    store i32 [[TMP129]], i32* [[M_CASTED45]], align 4
-// CHECK17-NEXT:    [[TMP130:%.*]] = load i32, i32* [[M_CASTED45]], align 4
-// CHECK17-NEXT:    [[TMP131:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP131]], i32* [[N_CASTED46]], align 4
-// CHECK17-NEXT:    [[TMP132:%.*]] = load i32, i32* [[N_CASTED46]], align 4
-// CHECK17-NEXT:    [[TMP133:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK17-NEXT:    [[TMP134:%.*]] = sext i32 [[TMP133]] to i64
-// CHECK17-NEXT:    [[TMP135:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES50]] to i8*
-// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP135]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
-// CHECK17-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP130]], i32* [[TMP137]], align 4
-// CHECK17-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP130]], i32* [[TMP139]], align 4
-// CHECK17-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP140]], align 4
-// CHECK17-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP132]], i32* [[TMP142]], align 4
-// CHECK17-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP132]], i32* [[TMP144]], align 4
-// CHECK17-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP145]], align 4
-// CHECK17-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP147]], align 4
-// CHECK17-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 2
-// CHECK17-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP149]], align 4
-// CHECK17-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 2
-// CHECK17-NEXT:    store i8* null, i8** [[TMP150]], align 4
-// CHECK17-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP152]], align 4
-// CHECK17-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 3
-// CHECK17-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i32**
-// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP154]], align 4
-// CHECK17-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES50]], i32 0, i32 3
-// CHECK17-NEXT:    store i64 [[TMP134]], i64* [[TMP155]], align 4
-// CHECK17-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 3
-// CHECK17-NEXT:    store i8* null, i8** [[TMP156]], align 4
-// CHECK17-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES50]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP160:%.*]] = load i32, i32* [[N]], align 4
-// CHECK17-NEXT:    store i32 [[TMP160]], i32* [[DOTCAPTURE_EXPR_52]], align 4
-// CHECK17-NEXT:    [[TMP161:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_52]], align 4
-// CHECK17-NEXT:    [[SUB54:%.*]] = sub nsw i32 [[TMP161]], 0
-// CHECK17-NEXT:    [[DIV55:%.*]] = sdiv i32 [[SUB54]], 1
-// CHECK17-NEXT:    [[SUB56:%.*]] = sub nsw i32 [[DIV55]], 1
-// CHECK17-NEXT:    store i32 [[SUB56]], i32* [[DOTCAPTURE_EXPR_53]], align 4
-// CHECK17-NEXT:    [[TMP162:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_53]], align 4
-// CHECK17-NEXT:    [[ADD57:%.*]] = add nsw i32 [[TMP162]], 1
-// CHECK17-NEXT:    [[TMP163:%.*]] = zext i32 [[ADD57]] to i64
-// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP163]])
-// CHECK17-NEXT:    [[TMP164:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174.region_id, i32 4, i8** [[TMP157]], i8** [[TMP158]], i64* [[TMP159]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP165:%.*]] = icmp ne i32 [[TMP164]], 0
-// CHECK17-NEXT:    br i1 [[TMP165]], label [[OMP_OFFLOAD_FAILED58:%.*]], label [[OMP_OFFLOAD_CONT59:%.*]]
-// CHECK17:       omp_offload.failed58:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174(i32 [[TMP130]], i32 [[TMP132]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT59]]
-// CHECK17:       omp_offload.cont59:
-// CHECK17-NEXT:    [[TMP166:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK17-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP166]])
+// CHECK17-NEXT:    store i32 [[TMP123]], i32* [[N_CASTED33]], align 4
+// CHECK17-NEXT:    [[TMP124:%.*]] = load i32, i32* [[N_CASTED33]], align 4
+// CHECK17-NEXT:    [[TMP125:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK17-NEXT:    [[TMP126:%.*]] = sext i32 [[TMP125]] to i64
+// CHECK17-NEXT:    [[TMP127:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES37]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP127]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
+// CHECK17-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP124]], i32* [[TMP129]], align 4
+// CHECK17-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP124]], i32* [[TMP131]], align 4
+// CHECK17-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i32 0, i32 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP132]], align 4
+// CHECK17-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP134]], align 4
+// CHECK17-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP136]], align 4
+// CHECK17-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i32 0, i32 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP137]], align 4
+// CHECK17-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP139]], align 4
+// CHECK17-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP141]], align 4
+// CHECK17-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES37]], i32 0, i32 2
+// CHECK17-NEXT:    store i64 [[TMP126]], i64* [[TMP142]], align 4
+// CHECK17-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i32 0, i32 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP143]], align 4
+// CHECK17-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES37]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP147:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP147]], i32* [[DOTCAPTURE_EXPR_39]], align 4
+// CHECK17-NEXT:    [[TMP148:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_39]], align 4
+// CHECK17-NEXT:    [[SUB41:%.*]] = sub nsw i32 [[TMP148]], 0
+// CHECK17-NEXT:    [[DIV42:%.*]] = sdiv i32 [[SUB41]], 1
+// CHECK17-NEXT:    [[SUB43:%.*]] = sub nsw i32 [[DIV42]], 1
+// CHECK17-NEXT:    store i32 [[SUB43]], i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK17-NEXT:    [[TMP149:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK17-NEXT:    [[ADD44:%.*]] = add nsw i32 [[TMP149]], 1
+// CHECK17-NEXT:    [[TMP150:%.*]] = zext i32 [[ADD44]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP150]])
+// CHECK17-NEXT:    [[KERNEL_ARGS45:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP151]], align 4
+// CHECK17-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP152]], align 4
+// CHECK17-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP144]], i8*** [[TMP153]], align 4
+// CHECK17-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP145]], i8*** [[TMP154]], align 4
+// CHECK17-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP146]], i64** [[TMP155]], align 4
+// CHECK17-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP156]], align 4
+// CHECK17-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP157]], align 4
+// CHECK17-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP158]], align 4
+// CHECK17-NEXT:    [[TMP159:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]])
+// CHECK17-NEXT:    [[TMP160:%.*]] = icmp ne i32 [[TMP159]], 0
+// CHECK17-NEXT:    br i1 [[TMP160]], label [[OMP_OFFLOAD_FAILED46:%.*]], label [[OMP_OFFLOAD_CONT47:%.*]]
+// CHECK17:       omp_offload.failed46:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169(i32 [[TMP124]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT47]]
+// CHECK17:       omp_offload.cont47:
+// CHECK17-NEXT:    [[TMP161:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    store i32 [[TMP161]], i32* [[M_CASTED48]], align 4
+// CHECK17-NEXT:    [[TMP162:%.*]] = load i32, i32* [[M_CASTED48]], align 4
+// CHECK17-NEXT:    [[TMP163:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP163]], i32* [[N_CASTED49]], align 4
+// CHECK17-NEXT:    [[TMP164:%.*]] = load i32, i32* [[N_CASTED49]], align 4
+// CHECK17-NEXT:    [[TMP165:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK17-NEXT:    [[TMP166:%.*]] = sext i32 [[TMP165]] to i64
+// CHECK17-NEXT:    [[TMP167:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES53]] to i8*
+// CHECK17-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP167]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
+// CHECK17-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP162]], i32* [[TMP169]], align 4
+// CHECK17-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP171:%.*]] = bitcast i8** [[TMP170]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP162]], i32* [[TMP171]], align 4
+// CHECK17-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP172]], align 4
+// CHECK17-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP174:%.*]] = bitcast i8** [[TMP173]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP164]], i32* [[TMP174]], align 4
+// CHECK17-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP176:%.*]] = bitcast i8** [[TMP175]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP164]], i32* [[TMP176]], align 4
+// CHECK17-NEXT:    [[TMP177:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP177]], align 4
+// CHECK17-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP179]], align 4
+// CHECK17-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 2
+// CHECK17-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP0]], i32* [[TMP181]], align 4
+// CHECK17-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 2
+// CHECK17-NEXT:    store i8* null, i8** [[TMP182]], align 4
+// CHECK17-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP184:%.*]] = bitcast i8** [[TMP183]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP184]], align 4
+// CHECK17-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 3
+// CHECK17-NEXT:    [[TMP186:%.*]] = bitcast i8** [[TMP185]] to i32**
+// CHECK17-NEXT:    store i32* [[VLA]], i32** [[TMP186]], align 4
+// CHECK17-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES53]], i32 0, i32 3
+// CHECK17-NEXT:    store i64 [[TMP166]], i64* [[TMP187]], align 4
+// CHECK17-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 3
+// CHECK17-NEXT:    store i8* null, i8** [[TMP188]], align 4
+// CHECK17-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES53]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP192:%.*]] = load i32, i32* [[N]], align 4
+// CHECK17-NEXT:    store i32 [[TMP192]], i32* [[DOTCAPTURE_EXPR_55]], align 4
+// CHECK17-NEXT:    [[TMP193:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_55]], align 4
+// CHECK17-NEXT:    [[SUB57:%.*]] = sub nsw i32 [[TMP193]], 0
+// CHECK17-NEXT:    [[DIV58:%.*]] = sdiv i32 [[SUB57]], 1
+// CHECK17-NEXT:    [[SUB59:%.*]] = sub nsw i32 [[DIV58]], 1
+// CHECK17-NEXT:    store i32 [[SUB59]], i32* [[DOTCAPTURE_EXPR_56]], align 4
+// CHECK17-NEXT:    [[TMP194:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_56]], align 4
+// CHECK17-NEXT:    [[ADD60:%.*]] = add nsw i32 [[TMP194]], 1
+// CHECK17-NEXT:    [[TMP195:%.*]] = zext i32 [[ADD60]] to i64
+// CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP195]])
+// CHECK17-NEXT:    [[KERNEL_ARGS61:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP196]], align 4
+// CHECK17-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 4, i32* [[TMP197]], align 4
+// CHECK17-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP189]], i8*** [[TMP198]], align 4
+// CHECK17-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP190]], i8*** [[TMP199]], align 4
+// CHECK17-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP191]], i64** [[TMP200]], align 4
+// CHECK17-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP201]], align 4
+// CHECK17-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP202]], align 4
+// CHECK17-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP203]], align 4
+// CHECK17-NEXT:    [[TMP204:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]])
+// CHECK17-NEXT:    [[TMP205:%.*]] = icmp ne i32 [[TMP204]], 0
+// CHECK17-NEXT:    br i1 [[TMP205]], label [[OMP_OFFLOAD_FAILED62:%.*]], label [[OMP_OFFLOAD_CONT63:%.*]]
+// CHECK17:       omp_offload.failed62:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174(i32 [[TMP162]], i32 [[TMP164]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT63]]
+// CHECK17:       omp_offload.cont63:
+// CHECK17-NEXT:    [[TMP206:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK17-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP206]])
 // CHECK17-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK17-NEXT:    [[TMP167:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP167]])
-// CHECK17-NEXT:    [[TMP168:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK17-NEXT:    ret i32 [[TMP168]]
+// CHECK17-NEXT:    [[TMP207:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP207]])
+// CHECK17-NEXT:    [[TMP208:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK17-NEXT:    ret i32 [[TMP208]]
 //
 //
 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154
@@ -11355,19 +12120,19 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 4
 // CHECK17-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    [[M_CASTED:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 4
-// CHECK17-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 4
-// CHECK17-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[M_CASTED19:%.*]] = alloca i32, align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [2 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [2 x i8*], align 4
-// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [2 x i8*], align 4
-// CHECK17-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 4
+// CHECK17-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 4
+// CHECK17-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[M_CASTED22:%.*]] = alloca i32, align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [2 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [2 x i8*], align 4
+// CHECK17-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [2 x i8*], align 4
+// CHECK17-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
 // CHECK17-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK17-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK17-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -11381,107 +12146,192 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l122.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK17-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK17-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK17-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK17-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK17-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK17-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK17-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l122.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK17-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l122([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
-// CHECK17-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK17-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK17-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK17-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK17-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK17-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK17-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l127.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK17-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK17:       omp_offload.failed5:
+// CHECK17-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK17-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK17-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK17-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l127.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK17-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK17-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK17:       omp_offload.failed6:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l127([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK17:       omp_offload.cont6:
-// CHECK17-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    store i32 [[TMP18]], i32* [[M_CASTED]], align 4
-// CHECK17-NEXT:    [[TMP19:%.*]] = load i32, i32* [[M_CASTED]], align 4
-// CHECK17-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP19]], i32* [[TMP21]], align 4
-// CHECK17-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP19]], i32* [[TMP23]], align 4
-// CHECK17-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP24]], align 4
-// CHECK17-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP26]], align 4
-// CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP28]], align 4
-// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP29]], align 4
-// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK17:       omp_offload.cont7:
+// CHECK17-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    store i32 [[TMP34]], i32* [[M_CASTED]], align 4
+// CHECK17-NEXT:    [[TMP35:%.*]] = load i32, i32* [[M_CASTED]], align 4
+// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP35]], i32* [[TMP37]], align 4
+// CHECK17-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP35]], i32* [[TMP39]], align 4
+// CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP40]], align 4
+// CHECK17-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP42]], align 4
+// CHECK17-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP44]], align 4
+// CHECK17-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP45]], align 4
+// CHECK17-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132.region_id, i32 2, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK17-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK17:       omp_offload.failed11:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132(i32 [[TMP19]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK17:       omp_offload.cont12:
-// CHECK17-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 4
-// CHECK17-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 4
-// CHECK17-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP38]], align 4
-// CHECK17-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK17-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 2, i32* [[TMP49]], align 4
+// CHECK17-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP46]], i8*** [[TMP50]], align 4
+// CHECK17-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 4
+// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP52]], align 4
+// CHECK17-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP53]], align 4
+// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP54]], align 4
+// CHECK17-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP55]], align 4
+// CHECK17-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK17-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK17-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK17:       omp_offload.failed13:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132(i32 [[TMP35]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK17:       omp_offload.cont14:
+// CHECK17-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP59]], align 4
+// CHECK17-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP61]], align 4
+// CHECK17-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP62]], align 4
+// CHECK17-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP41:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l137.region_id, i32 1, i8** [[TMP39]], i8** [[TMP40]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP42:%.*]] = icmp ne i32 [[TMP41]], 0
-// CHECK17-NEXT:    br i1 [[TMP42]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK17:       omp_offload.failed17:
+// CHECK17-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP65]], align 4
+// CHECK17-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK17-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP63]], i8*** [[TMP67]], align 4
+// CHECK17-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK17-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP69]], align 4
+// CHECK17-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK17-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP71]], align 4
+// CHECK17-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK17-NEXT:    [[TMP73:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l137.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK17-NEXT:    [[TMP74:%.*]] = icmp ne i32 [[TMP73]], 0
+// CHECK17-NEXT:    br i1 [[TMP74]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK17:       omp_offload.failed20:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l137([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK17:       omp_offload.cont18:
-// CHECK17-NEXT:    [[TMP43:%.*]] = load i32, i32* [[M]], align 4
-// CHECK17-NEXT:    store i32 [[TMP43]], i32* [[M_CASTED19]], align 4
-// CHECK17-NEXT:    [[TMP44:%.*]] = load i32, i32* [[M_CASTED19]], align 4
-// CHECK17-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP44]], i32* [[TMP46]], align 4
-// CHECK17-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
-// CHECK17-NEXT:    store i32 [[TMP44]], i32* [[TMP48]], align 4
-// CHECK17-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
-// CHECK17-NEXT:    store i8* null, i8** [[TMP49]], align 4
-// CHECK17-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP51]], align 4
-// CHECK17-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK17-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to [10 x i32]**
-// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP53]], align 4
-// CHECK17-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
-// CHECK17-NEXT:    store i8* null, i8** [[TMP54]], align 4
-// CHECK17-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK17:       omp_offload.cont21:
+// CHECK17-NEXT:    [[TMP75:%.*]] = load i32, i32* [[M]], align 4
+// CHECK17-NEXT:    store i32 [[TMP75]], i32* [[M_CASTED22]], align 4
+// CHECK17-NEXT:    [[TMP76:%.*]] = load i32, i32* [[M_CASTED22]], align 4
+// CHECK17-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP76]], i32* [[TMP78]], align 4
+// CHECK17-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32*
+// CHECK17-NEXT:    store i32 [[TMP76]], i32* [[TMP80]], align 4
+// CHECK17-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 0
+// CHECK17-NEXT:    store i8* null, i8** [[TMP81]], align 4
+// CHECK17-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP83]], align 4
+// CHECK17-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
+// CHECK17-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to [10 x i32]**
+// CHECK17-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP85]], align 4
+// CHECK17-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 1
+// CHECK17-NEXT:    store i8* null, i8** [[TMP86]], align 4
+// CHECK17-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK17-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK17-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
-// CHECK17-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
-// CHECK17:       omp_offload.failed24:
-// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142(i32 [[TMP44]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
-// CHECK17:       omp_offload.cont25:
+// CHECK17-NEXT:    [[KERNEL_ARGS27:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK17-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP89]], align 4
+// CHECK17-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 2, i32* [[TMP90]], align 4
+// CHECK17-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP87]], i8*** [[TMP91]], align 4
+// CHECK17-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP88]], i8*** [[TMP92]], align 4
+// CHECK17-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP93]], align 4
+// CHECK17-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP94]], align 4
+// CHECK17-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP95]], align 4
+// CHECK17-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP96]], align 4
+// CHECK17-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]])
+// CHECK17-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
+// CHECK17-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
+// CHECK17:       omp_offload.failed28:
+// CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142(i32 [[TMP76]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
+// CHECK17:       omp_offload.cont29:
 // CHECK17-NEXT:    ret i32 0
 //
 //
@@ -12365,31 +13215,31 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[M_CASTED:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[N_CASTED17:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS18:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS19:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS20:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES21:%.*]] = alloca [4 x i64], align 4
-// CHECK19-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_23:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [4 x i64], align 4
+// CHECK19-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[N_CASTED31:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS32:%.*]] = alloca [3 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS33:%.*]] = alloca [3 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS34:%.*]] = alloca [3 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES35:%.*]] = alloca [3 x i64], align 4
-// CHECK19-NEXT:    [[_TMP36:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_37:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_38:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[M_CASTED45:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[N_CASTED46:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS47:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS48:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS49:%.*]] = alloca [4 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES50:%.*]] = alloca [4 x i64], align 4
-// CHECK19-NEXT:    [[_TMP51:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_52:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_53:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[N_CASTED33:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS34:%.*]] = alloca [3 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS35:%.*]] = alloca [3 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS36:%.*]] = alloca [3 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES37:%.*]] = alloca [3 x i64], align 4
+// CHECK19-NEXT:    [[_TMP38:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_39:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_40:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[M_CASTED48:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[N_CASTED49:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS50:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS51:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS52:%.*]] = alloca [4 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_SIZES53:%.*]] = alloca [4 x i64], align 4
+// CHECK19-NEXT:    [[_TMP54:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_55:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTCAPTURE_EXPR_56:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK19-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK19-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -12447,258 +13297,343 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK19-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK19-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK19-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK19-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK19-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK19-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK19-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK19-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK19-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
-// CHECK19-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK19-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK19-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK19-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
-// CHECK19-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK19-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK19-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK19-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK19-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK19-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK19-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK19-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK19-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK19-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK19-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK19-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK19-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.4 to i8*), i32 24, i1 false)
+// CHECK19-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK19-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK19-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK19-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK19-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK19-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK19-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK19-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK19-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK19-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK19-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK19-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK19-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK19-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK19-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP59]])
-// CHECK19-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK19-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK19:       omp_offload.failed15:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK19:       omp_offload.cont16:
-// CHECK19-NEXT:    [[TMP62:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP62]], i32* [[M_CASTED]], align 4
-// CHECK19-NEXT:    [[TMP63:%.*]] = load i32, i32* [[M_CASTED]], align 4
-// CHECK19-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP64]], i32* [[N_CASTED17]], align 4
-// CHECK19-NEXT:    [[TMP65:%.*]] = load i32, i32* [[N_CASTED17]], align 4
-// CHECK19-NEXT:    [[TMP66:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP67:%.*]] = sext i32 [[TMP66]] to i64
-// CHECK19-NEXT:    [[TMP68:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES21]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP68]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
-// CHECK19-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP63]], i32* [[TMP70]], align 4
-// CHECK19-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP72:%.*]] = bitcast i8** [[TMP71]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP63]], i32* [[TMP72]], align 4
-// CHECK19-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP73]], align 4
-// CHECK19-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP65]], i32* [[TMP75]], align 4
-// CHECK19-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP77:%.*]] = bitcast i8** [[TMP76]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP65]], i32* [[TMP77]], align 4
-// CHECK19-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP78]], align 4
-// CHECK19-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP80]], align 4
-// CHECK19-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP82]], align 4
-// CHECK19-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP83]], align 4
-// CHECK19-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP85]], align 4
-// CHECK19-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP87]], align 4
-// CHECK19-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 3
-// CHECK19-NEXT:    store i64 [[TMP67]], i64* [[TMP88]], align 4
-// CHECK19-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK19-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK19-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP67]])
+// CHECK19-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK19-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK19-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK19-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK19-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK19-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK19-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK19-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK19-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK19-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK19-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK19:       omp_offload.failed16:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l159(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK19:       omp_offload.cont17:
+// CHECK19-NEXT:    [[TMP78:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP78]], i32* [[M_CASTED]], align 4
+// CHECK19-NEXT:    [[TMP79:%.*]] = load i32, i32* [[M_CASTED]], align 4
+// CHECK19-NEXT:    [[TMP80:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP80]], i32* [[N_CASTED18]], align 4
+// CHECK19-NEXT:    [[TMP81:%.*]] = load i32, i32* [[N_CASTED18]], align 4
+// CHECK19-NEXT:    [[TMP82:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP83:%.*]] = sext i32 [[TMP82]] to i64
+// CHECK19-NEXT:    [[TMP84:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP84]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.8 to i8*), i32 32, i1 false)
+// CHECK19-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP79]], i32* [[TMP86]], align 4
+// CHECK19-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP88:%.*]] = bitcast i8** [[TMP87]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP79]], i32* [[TMP88]], align 4
+// CHECK19-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
 // CHECK19-NEXT:    store i8* null, i8** [[TMP89]], align 4
-// CHECK19-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP93:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP93]], i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK19-NEXT:    [[TMP94:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK19-NEXT:    [[SUB25:%.*]] = sub nsw i32 [[TMP94]], 0
-// CHECK19-NEXT:    [[DIV26:%.*]] = sdiv i32 [[SUB25]], 1
-// CHECK19-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[DIV26]], 1
-// CHECK19-NEXT:    store i32 [[SUB27]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK19-NEXT:    [[TMP95:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK19-NEXT:    [[ADD28:%.*]] = add nsw i32 [[TMP95]], 1
-// CHECK19-NEXT:    [[TMP96:%.*]] = zext i32 [[ADD28]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP96]])
-// CHECK19-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164.region_id, i32 4, i8** [[TMP90]], i8** [[TMP91]], i64* [[TMP92]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
-// CHECK19-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
-// CHECK19:       omp_offload.failed29:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164(i32 [[TMP63]], i32 [[TMP65]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
-// CHECK19:       omp_offload.cont30:
-// CHECK19-NEXT:    [[TMP99:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP99]], i32* [[N_CASTED31]], align 4
-// CHECK19-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N_CASTED31]], align 4
-// CHECK19-NEXT:    [[TMP101:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP102:%.*]] = sext i32 [[TMP101]] to i64
-// CHECK19-NEXT:    [[TMP103:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES35]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP103]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
-// CHECK19-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP105:%.*]] = bitcast i8** [[TMP104]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP100]], i32* [[TMP105]], align 4
-// CHECK19-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP107:%.*]] = bitcast i8** [[TMP106]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP100]], i32* [[TMP107]], align 4
-// CHECK19-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP108]], align 4
-// CHECK19-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP110:%.*]] = bitcast i8** [[TMP109]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP110]], align 4
-// CHECK19-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP112:%.*]] = bitcast i8** [[TMP111]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP112]], align 4
-// CHECK19-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP113]], align 4
-// CHECK19-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP115:%.*]] = bitcast i8** [[TMP114]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP115]], align 4
-// CHECK19-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP117:%.*]] = bitcast i8** [[TMP116]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP117]], align 4
-// CHECK19-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES35]], i32 0, i32 2
-// CHECK19-NEXT:    store i64 [[TMP102]], i64* [[TMP118]], align 4
-// CHECK19-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS34]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP119]], align 4
-// CHECK19-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS32]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP121:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS33]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP122:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES35]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP81]], i32* [[TMP91]], align 4
+// CHECK19-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP93:%.*]] = bitcast i8** [[TMP92]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP81]], i32* [[TMP93]], align 4
+// CHECK19-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP94]], align 4
+// CHECK19-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP96]], align 4
+// CHECK19-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP98:%.*]] = bitcast i8** [[TMP97]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP98]], align 4
+// CHECK19-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP99]], align 4
+// CHECK19-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP101:%.*]] = bitcast i8** [[TMP100]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP101]], align 4
+// CHECK19-NEXT:    [[TMP102:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP103:%.*]] = bitcast i8** [[TMP102]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP103]], align 4
+// CHECK19-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 3
+// CHECK19-NEXT:    store i64 [[TMP83]], i64* [[TMP104]], align 4
+// CHECK19-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 3
+// CHECK19-NEXT:    store i8* null, i8** [[TMP105]], align 4
+// CHECK19-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP109:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP109]], i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK19-NEXT:    [[TMP110:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK19-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP110]], 0
+// CHECK19-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
+// CHECK19-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
+// CHECK19-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK19-NEXT:    [[TMP111:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK19-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP111]], 1
+// CHECK19-NEXT:    [[TMP112:%.*]] = zext i32 [[ADD29]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP112]])
+// CHECK19-NEXT:    [[KERNEL_ARGS30:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP113]], align 4
+// CHECK19-NEXT:    [[TMP114:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 4, i32* [[TMP114]], align 4
+// CHECK19-NEXT:    [[TMP115:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP106]], i8*** [[TMP115]], align 4
+// CHECK19-NEXT:    [[TMP116:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP107]], i8*** [[TMP116]], align 4
+// CHECK19-NEXT:    [[TMP117:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP108]], i64** [[TMP117]], align 4
+// CHECK19-NEXT:    [[TMP118:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP118]], align 4
+// CHECK19-NEXT:    [[TMP119:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP119]], align 4
+// CHECK19-NEXT:    [[TMP120:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP120]], align 4
+// CHECK19-NEXT:    [[TMP121:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]])
+// CHECK19-NEXT:    [[TMP122:%.*]] = icmp ne i32 [[TMP121]], 0
+// CHECK19-NEXT:    br i1 [[TMP122]], label [[OMP_OFFLOAD_FAILED31:%.*]], label [[OMP_OFFLOAD_CONT32:%.*]]
+// CHECK19:       omp_offload.failed31:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l164(i32 [[TMP79]], i32 [[TMP81]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT32]]
+// CHECK19:       omp_offload.cont32:
 // CHECK19-NEXT:    [[TMP123:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP123]], i32* [[DOTCAPTURE_EXPR_37]], align 4
-// CHECK19-NEXT:    [[TMP124:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_37]], align 4
-// CHECK19-NEXT:    [[SUB39:%.*]] = sub nsw i32 [[TMP124]], 0
-// CHECK19-NEXT:    [[DIV40:%.*]] = sdiv i32 [[SUB39]], 1
-// CHECK19-NEXT:    [[SUB41:%.*]] = sub nsw i32 [[DIV40]], 1
-// CHECK19-NEXT:    store i32 [[SUB41]], i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK19-NEXT:    [[TMP125:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_38]], align 4
-// CHECK19-NEXT:    [[ADD42:%.*]] = add nsw i32 [[TMP125]], 1
-// CHECK19-NEXT:    [[TMP126:%.*]] = zext i32 [[ADD42]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP126]])
-// CHECK19-NEXT:    [[TMP127:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169.region_id, i32 3, i8** [[TMP120]], i8** [[TMP121]], i64* [[TMP122]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP128:%.*]] = icmp ne i32 [[TMP127]], 0
-// CHECK19-NEXT:    br i1 [[TMP128]], label [[OMP_OFFLOAD_FAILED43:%.*]], label [[OMP_OFFLOAD_CONT44:%.*]]
-// CHECK19:       omp_offload.failed43:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169(i32 [[TMP100]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT44]]
-// CHECK19:       omp_offload.cont44:
-// CHECK19-NEXT:    [[TMP129:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP129]], i32* [[M_CASTED45]], align 4
-// CHECK19-NEXT:    [[TMP130:%.*]] = load i32, i32* [[M_CASTED45]], align 4
-// CHECK19-NEXT:    [[TMP131:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP131]], i32* [[N_CASTED46]], align 4
-// CHECK19-NEXT:    [[TMP132:%.*]] = load i32, i32* [[N_CASTED46]], align 4
-// CHECK19-NEXT:    [[TMP133:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK19-NEXT:    [[TMP134:%.*]] = sext i32 [[TMP133]] to i64
-// CHECK19-NEXT:    [[TMP135:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES50]] to i8*
-// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP135]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
-// CHECK19-NEXT:    [[TMP136:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP137:%.*]] = bitcast i8** [[TMP136]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP130]], i32* [[TMP137]], align 4
-// CHECK19-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP130]], i32* [[TMP139]], align 4
-// CHECK19-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP140]], align 4
-// CHECK19-NEXT:    [[TMP141:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP142:%.*]] = bitcast i8** [[TMP141]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP132]], i32* [[TMP142]], align 4
-// CHECK19-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP144:%.*]] = bitcast i8** [[TMP143]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP132]], i32* [[TMP144]], align 4
-// CHECK19-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP145]], align 4
-// CHECK19-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP147:%.*]] = bitcast i8** [[TMP146]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP147]], align 4
-// CHECK19-NEXT:    [[TMP148:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 2
-// CHECK19-NEXT:    [[TMP149:%.*]] = bitcast i8** [[TMP148]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP149]], align 4
-// CHECK19-NEXT:    [[TMP150:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 2
-// CHECK19-NEXT:    store i8* null, i8** [[TMP150]], align 4
-// CHECK19-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP152:%.*]] = bitcast i8** [[TMP151]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP152]], align 4
-// CHECK19-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 3
-// CHECK19-NEXT:    [[TMP154:%.*]] = bitcast i8** [[TMP153]] to i32**
-// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP154]], align 4
-// CHECK19-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES50]], i32 0, i32 3
-// CHECK19-NEXT:    store i64 [[TMP134]], i64* [[TMP155]], align 4
-// CHECK19-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS49]], i32 0, i32 3
-// CHECK19-NEXT:    store i8* null, i8** [[TMP156]], align 4
-// CHECK19-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS47]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS48]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP159:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES50]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP160:%.*]] = load i32, i32* [[N]], align 4
-// CHECK19-NEXT:    store i32 [[TMP160]], i32* [[DOTCAPTURE_EXPR_52]], align 4
-// CHECK19-NEXT:    [[TMP161:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_52]], align 4
-// CHECK19-NEXT:    [[SUB54:%.*]] = sub nsw i32 [[TMP161]], 0
-// CHECK19-NEXT:    [[DIV55:%.*]] = sdiv i32 [[SUB54]], 1
-// CHECK19-NEXT:    [[SUB56:%.*]] = sub nsw i32 [[DIV55]], 1
-// CHECK19-NEXT:    store i32 [[SUB56]], i32* [[DOTCAPTURE_EXPR_53]], align 4
-// CHECK19-NEXT:    [[TMP162:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_53]], align 4
-// CHECK19-NEXT:    [[ADD57:%.*]] = add nsw i32 [[TMP162]], 1
-// CHECK19-NEXT:    [[TMP163:%.*]] = zext i32 [[ADD57]] to i64
-// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP163]])
-// CHECK19-NEXT:    [[TMP164:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174.region_id, i32 4, i8** [[TMP157]], i8** [[TMP158]], i64* [[TMP159]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP165:%.*]] = icmp ne i32 [[TMP164]], 0
-// CHECK19-NEXT:    br i1 [[TMP165]], label [[OMP_OFFLOAD_FAILED58:%.*]], label [[OMP_OFFLOAD_CONT59:%.*]]
-// CHECK19:       omp_offload.failed58:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174(i32 [[TMP130]], i32 [[TMP132]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT59]]
-// CHECK19:       omp_offload.cont59:
-// CHECK19-NEXT:    [[TMP166:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK19-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP166]])
+// CHECK19-NEXT:    store i32 [[TMP123]], i32* [[N_CASTED33]], align 4
+// CHECK19-NEXT:    [[TMP124:%.*]] = load i32, i32* [[N_CASTED33]], align 4
+// CHECK19-NEXT:    [[TMP125:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP126:%.*]] = sext i32 [[TMP125]] to i64
+// CHECK19-NEXT:    [[TMP127:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES37]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP127]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.12 to i8*), i32 24, i1 false)
+// CHECK19-NEXT:    [[TMP128:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP129:%.*]] = bitcast i8** [[TMP128]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP124]], i32* [[TMP129]], align 4
+// CHECK19-NEXT:    [[TMP130:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP131:%.*]] = bitcast i8** [[TMP130]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP124]], i32* [[TMP131]], align 4
+// CHECK19-NEXT:    [[TMP132:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP132]], align 4
+// CHECK19-NEXT:    [[TMP133:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP134:%.*]] = bitcast i8** [[TMP133]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP134]], align 4
+// CHECK19-NEXT:    [[TMP135:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP136:%.*]] = bitcast i8** [[TMP135]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP136]], align 4
+// CHECK19-NEXT:    [[TMP137:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP137]], align 4
+// CHECK19-NEXT:    [[TMP138:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP139:%.*]] = bitcast i8** [[TMP138]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP139]], align 4
+// CHECK19-NEXT:    [[TMP140:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP141:%.*]] = bitcast i8** [[TMP140]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP141]], align 4
+// CHECK19-NEXT:    [[TMP142:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES37]], i32 0, i32 2
+// CHECK19-NEXT:    store i64 [[TMP126]], i64* [[TMP142]], align 4
+// CHECK19-NEXT:    [[TMP143:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS36]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP143]], align 4
+// CHECK19-NEXT:    [[TMP144:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS34]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP145:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS35]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP146:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES37]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP147:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP147]], i32* [[DOTCAPTURE_EXPR_39]], align 4
+// CHECK19-NEXT:    [[TMP148:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_39]], align 4
+// CHECK19-NEXT:    [[SUB41:%.*]] = sub nsw i32 [[TMP148]], 0
+// CHECK19-NEXT:    [[DIV42:%.*]] = sdiv i32 [[SUB41]], 1
+// CHECK19-NEXT:    [[SUB43:%.*]] = sub nsw i32 [[DIV42]], 1
+// CHECK19-NEXT:    store i32 [[SUB43]], i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK19-NEXT:    [[TMP149:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_40]], align 4
+// CHECK19-NEXT:    [[ADD44:%.*]] = add nsw i32 [[TMP149]], 1
+// CHECK19-NEXT:    [[TMP150:%.*]] = zext i32 [[ADD44]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP150]])
+// CHECK19-NEXT:    [[KERNEL_ARGS45:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP151:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP151]], align 4
+// CHECK19-NEXT:    [[TMP152:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP152]], align 4
+// CHECK19-NEXT:    [[TMP153:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP144]], i8*** [[TMP153]], align 4
+// CHECK19-NEXT:    [[TMP154:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP145]], i8*** [[TMP154]], align 4
+// CHECK19-NEXT:    [[TMP155:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP146]], i64** [[TMP155]], align 4
+// CHECK19-NEXT:    [[TMP156:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.13, i32 0, i32 0), i64** [[TMP156]], align 4
+// CHECK19-NEXT:    [[TMP157:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP157]], align 4
+// CHECK19-NEXT:    [[TMP158:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP158]], align 4
+// CHECK19-NEXT:    [[TMP159:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS45]])
+// CHECK19-NEXT:    [[TMP160:%.*]] = icmp ne i32 [[TMP159]], 0
+// CHECK19-NEXT:    br i1 [[TMP160]], label [[OMP_OFFLOAD_FAILED46:%.*]], label [[OMP_OFFLOAD_CONT47:%.*]]
+// CHECK19:       omp_offload.failed46:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l169(i32 [[TMP124]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT47]]
+// CHECK19:       omp_offload.cont47:
+// CHECK19-NEXT:    [[TMP161:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP161]], i32* [[M_CASTED48]], align 4
+// CHECK19-NEXT:    [[TMP162:%.*]] = load i32, i32* [[M_CASTED48]], align 4
+// CHECK19-NEXT:    [[TMP163:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP163]], i32* [[N_CASTED49]], align 4
+// CHECK19-NEXT:    [[TMP164:%.*]] = load i32, i32* [[N_CASTED49]], align 4
+// CHECK19-NEXT:    [[TMP165:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK19-NEXT:    [[TMP166:%.*]] = sext i32 [[TMP165]] to i64
+// CHECK19-NEXT:    [[TMP167:%.*]] = bitcast [4 x i64]* [[DOTOFFLOAD_SIZES53]] to i8*
+// CHECK19-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP167]], i8* align 4 bitcast ([4 x i64]* @.offload_sizes.16 to i8*), i32 32, i1 false)
+// CHECK19-NEXT:    [[TMP168:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP169:%.*]] = bitcast i8** [[TMP168]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP162]], i32* [[TMP169]], align 4
+// CHECK19-NEXT:    [[TMP170:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP171:%.*]] = bitcast i8** [[TMP170]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP162]], i32* [[TMP171]], align 4
+// CHECK19-NEXT:    [[TMP172:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP172]], align 4
+// CHECK19-NEXT:    [[TMP173:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP174:%.*]] = bitcast i8** [[TMP173]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP164]], i32* [[TMP174]], align 4
+// CHECK19-NEXT:    [[TMP175:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP176:%.*]] = bitcast i8** [[TMP175]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP164]], i32* [[TMP176]], align 4
+// CHECK19-NEXT:    [[TMP177:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP177]], align 4
+// CHECK19-NEXT:    [[TMP178:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP179:%.*]] = bitcast i8** [[TMP178]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP179]], align 4
+// CHECK19-NEXT:    [[TMP180:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 2
+// CHECK19-NEXT:    [[TMP181:%.*]] = bitcast i8** [[TMP180]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP0]], i32* [[TMP181]], align 4
+// CHECK19-NEXT:    [[TMP182:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 2
+// CHECK19-NEXT:    store i8* null, i8** [[TMP182]], align 4
+// CHECK19-NEXT:    [[TMP183:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP184:%.*]] = bitcast i8** [[TMP183]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP184]], align 4
+// CHECK19-NEXT:    [[TMP185:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 3
+// CHECK19-NEXT:    [[TMP186:%.*]] = bitcast i8** [[TMP185]] to i32**
+// CHECK19-NEXT:    store i32* [[VLA]], i32** [[TMP186]], align 4
+// CHECK19-NEXT:    [[TMP187:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES53]], i32 0, i32 3
+// CHECK19-NEXT:    store i64 [[TMP166]], i64* [[TMP187]], align 4
+// CHECK19-NEXT:    [[TMP188:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_MAPPERS52]], i32 0, i32 3
+// CHECK19-NEXT:    store i8* null, i8** [[TMP188]], align 4
+// CHECK19-NEXT:    [[TMP189:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS50]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP190:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS51]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP191:%.*]] = getelementptr inbounds [4 x i64], [4 x i64]* [[DOTOFFLOAD_SIZES53]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP192:%.*]] = load i32, i32* [[N]], align 4
+// CHECK19-NEXT:    store i32 [[TMP192]], i32* [[DOTCAPTURE_EXPR_55]], align 4
+// CHECK19-NEXT:    [[TMP193:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_55]], align 4
+// CHECK19-NEXT:    [[SUB57:%.*]] = sub nsw i32 [[TMP193]], 0
+// CHECK19-NEXT:    [[DIV58:%.*]] = sdiv i32 [[SUB57]], 1
+// CHECK19-NEXT:    [[SUB59:%.*]] = sub nsw i32 [[DIV58]], 1
+// CHECK19-NEXT:    store i32 [[SUB59]], i32* [[DOTCAPTURE_EXPR_56]], align 4
+// CHECK19-NEXT:    [[TMP194:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_56]], align 4
+// CHECK19-NEXT:    [[ADD60:%.*]] = add nsw i32 [[TMP194]], 1
+// CHECK19-NEXT:    [[TMP195:%.*]] = zext i32 [[ADD60]] to i64
+// CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 [[TMP195]])
+// CHECK19-NEXT:    [[KERNEL_ARGS61:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP196:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP196]], align 4
+// CHECK19-NEXT:    [[TMP197:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 4, i32* [[TMP197]], align 4
+// CHECK19-NEXT:    [[TMP198:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP189]], i8*** [[TMP198]], align 4
+// CHECK19-NEXT:    [[TMP199:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP190]], i8*** [[TMP199]], align 4
+// CHECK19-NEXT:    [[TMP200:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP191]], i64** [[TMP200]], align 4
+// CHECK19-NEXT:    [[TMP201:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.17, i32 0, i32 0), i64** [[TMP201]], align 4
+// CHECK19-NEXT:    [[TMP202:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP202]], align 4
+// CHECK19-NEXT:    [[TMP203:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP203]], align 4
+// CHECK19-NEXT:    [[TMP204:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS61]])
+// CHECK19-NEXT:    [[TMP205:%.*]] = icmp ne i32 [[TMP204]], 0
+// CHECK19-NEXT:    br i1 [[TMP205]], label [[OMP_OFFLOAD_FAILED62:%.*]], label [[OMP_OFFLOAD_CONT63:%.*]]
+// CHECK19:       omp_offload.failed62:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l174(i32 [[TMP162]], i32 [[TMP164]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT63]]
+// CHECK19:       omp_offload.cont63:
+// CHECK19-NEXT:    [[TMP206:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK19-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP206]])
 // CHECK19-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK19-NEXT:    [[TMP167:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP167]])
-// CHECK19-NEXT:    [[TMP168:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK19-NEXT:    ret i32 [[TMP168]]
+// CHECK19-NEXT:    [[TMP207:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP207]])
+// CHECK19-NEXT:    [[TMP208:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK19-NEXT:    ret i32 [[TMP208]]
 //
 //
 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l154
@@ -13918,19 +14853,19 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 4
 // CHECK19-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    [[M_CASTED:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS13:%.*]] = alloca [1 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS14:%.*]] = alloca [1 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS15:%.*]] = alloca [1 x i8*], align 4
-// CHECK19-NEXT:    [[_TMP16:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[M_CASTED19:%.*]] = alloca i32, align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [2 x i8*], align 4
-// CHECK19-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS15:%.*]] = alloca [1 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS16:%.*]] = alloca [1 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS17:%.*]] = alloca [1 x i8*], align 4
+// CHECK19-NEXT:    [[_TMP18:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[M_CASTED22:%.*]] = alloca i32, align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_BASEPTRS23:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_PTRS24:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[DOTOFFLOAD_MAPPERS25:%.*]] = alloca [2 x i8*], align 4
+// CHECK19-NEXT:    [[_TMP26:%.*]] = alloca i32, align 4
 // CHECK19-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK19-NEXT:    store i32 10, i32* [[M]], align 4
 // CHECK19-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
@@ -13944,107 +14879,192 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l122.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK19-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK19-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK19-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK19-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.20, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK19-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.21, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK19-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK19-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l122.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK19-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l122([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
-// CHECK19-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK19-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK19-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK19-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK19-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK19-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l127.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK19-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK19:       omp_offload.failed5:
+// CHECK19-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK19-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.24, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.25, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK19-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l127.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK19-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK19-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK19:       omp_offload.failed6:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l127([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK19:       omp_offload.cont6:
-// CHECK19-NEXT:    [[TMP18:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP18]], i32* [[M_CASTED]], align 4
-// CHECK19-NEXT:    [[TMP19:%.*]] = load i32, i32* [[M_CASTED]], align 4
-// CHECK19-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP19]], i32* [[TMP21]], align 4
-// CHECK19-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP23:%.*]] = bitcast i8** [[TMP22]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP19]], i32* [[TMP23]], align 4
-// CHECK19-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP24]], align 4
-// CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP26]], align 4
-// CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP28]], align 4
-// CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP29]], align 4
-// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK19:       omp_offload.cont7:
+// CHECK19-NEXT:    [[TMP34:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP34]], i32* [[M_CASTED]], align 4
+// CHECK19-NEXT:    [[TMP35:%.*]] = load i32, i32* [[M_CASTED]], align 4
+// CHECK19-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP35]], i32* [[TMP37]], align 4
+// CHECK19-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP35]], i32* [[TMP39]], align 4
+// CHECK19-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP40]], align 4
+// CHECK19-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP42]], align 4
+// CHECK19-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP44]], align 4
+// CHECK19-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP45]], align 4
+// CHECK19-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132.region_id, i32 2, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK19-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK19:       omp_offload.failed11:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132(i32 [[TMP19]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK19:       omp_offload.cont12:
-// CHECK19-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 4
-// CHECK19-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 4
-// CHECK19-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS15]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP38]], align 4
-// CHECK19-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS13]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS14]], i32 0, i32 0
+// CHECK19-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP48]], align 4
+// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 2, i32* [[TMP49]], align 4
+// CHECK19-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP46]], i8*** [[TMP50]], align 4
+// CHECK19-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP47]], i8*** [[TMP51]], align 4
+// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.28, i32 0, i32 0), i64** [[TMP52]], align 4
+// CHECK19-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.29, i32 0, i32 0), i64** [[TMP53]], align 4
+// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP54]], align 4
+// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP55]], align 4
+// CHECK19-NEXT:    [[TMP56:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK19-NEXT:    [[TMP57:%.*]] = icmp ne i32 [[TMP56]], 0
+// CHECK19-NEXT:    br i1 [[TMP57]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK19:       omp_offload.failed13:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l132(i32 [[TMP35]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK19:       omp_offload.cont14:
+// CHECK19-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP59]], align 4
+// CHECK19-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP61]], align 4
+// CHECK19-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS17]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP62]], align 4
+// CHECK19-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS15]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS16]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP41:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l137.region_id, i32 1, i8** [[TMP39]], i8** [[TMP40]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP42:%.*]] = icmp ne i32 [[TMP41]], 0
-// CHECK19-NEXT:    br i1 [[TMP42]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
-// CHECK19:       omp_offload.failed17:
+// CHECK19-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP65]], align 4
+// CHECK19-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 1, i32* [[TMP66]], align 4
+// CHECK19-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP63]], i8*** [[TMP67]], align 4
+// CHECK19-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP64]], i8*** [[TMP68]], align 4
+// CHECK19-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.32, i32 0, i32 0), i64** [[TMP69]], align 4
+// CHECK19-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.33, i32 0, i32 0), i64** [[TMP70]], align 4
+// CHECK19-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP71]], align 4
+// CHECK19-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP72]], align 4
+// CHECK19-NEXT:    [[TMP73:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l137.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK19-NEXT:    [[TMP74:%.*]] = icmp ne i32 [[TMP73]], 0
+// CHECK19-NEXT:    br i1 [[TMP74]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK19:       omp_offload.failed20:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l137([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
-// CHECK19:       omp_offload.cont18:
-// CHECK19-NEXT:    [[TMP43:%.*]] = load i32, i32* [[M]], align 4
-// CHECK19-NEXT:    store i32 [[TMP43]], i32* [[M_CASTED19]], align 4
-// CHECK19-NEXT:    [[TMP44:%.*]] = load i32, i32* [[M_CASTED19]], align 4
-// CHECK19-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP44]], i32* [[TMP46]], align 4
-// CHECK19-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
-// CHECK19-NEXT:    store i32 [[TMP44]], i32* [[TMP48]], align 4
-// CHECK19-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 0
-// CHECK19-NEXT:    store i8* null, i8** [[TMP49]], align 4
-// CHECK19-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP51]], align 4
-// CHECK19-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK19-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to [10 x i32]**
-// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP53]], align 4
-// CHECK19-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i32 0, i32 1
-// CHECK19-NEXT:    store i8* null, i8** [[TMP54]], align 4
-// CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK19:       omp_offload.cont21:
+// CHECK19-NEXT:    [[TMP75:%.*]] = load i32, i32* [[M]], align 4
+// CHECK19-NEXT:    store i32 [[TMP75]], i32* [[M_CASTED22]], align 4
+// CHECK19-NEXT:    [[TMP76:%.*]] = load i32, i32* [[M_CASTED22]], align 4
+// CHECK19-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP76]], i32* [[TMP78]], align 4
+// CHECK19-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32*
+// CHECK19-NEXT:    store i32 [[TMP76]], i32* [[TMP80]], align 4
+// CHECK19-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 0
+// CHECK19-NEXT:    store i8* null, i8** [[TMP81]], align 4
+// CHECK19-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP83:%.*]] = bitcast i8** [[TMP82]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP83]], align 4
+// CHECK19-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 1
+// CHECK19-NEXT:    [[TMP85:%.*]] = bitcast i8** [[TMP84]] to [10 x i32]**
+// CHECK19-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP85]], align 4
+// CHECK19-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS25]], i32 0, i32 1
+// CHECK19-NEXT:    store i8* null, i8** [[TMP86]], align 4
+// CHECK19-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS23]], i32 0, i32 0
+// CHECK19-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS24]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 10)
-// CHECK19-NEXT:    [[TMP57:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP58:%.*]] = icmp ne i32 [[TMP57]], 0
-// CHECK19-NEXT:    br i1 [[TMP58]], label [[OMP_OFFLOAD_FAILED24:%.*]], label [[OMP_OFFLOAD_CONT25:%.*]]
-// CHECK19:       omp_offload.failed24:
-// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142(i32 [[TMP44]], [10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT25]]
-// CHECK19:       omp_offload.cont25:
+// CHECK19-NEXT:    [[KERNEL_ARGS27:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK19-NEXT:    [[TMP89:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP89]], align 4
+// CHECK19-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 2, i32* [[TMP90]], align 4
+// CHECK19-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP87]], i8*** [[TMP91]], align 4
+// CHECK19-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP88]], i8*** [[TMP92]], align 4
+// CHECK19-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.36, i32 0, i32 0), i64** [[TMP93]], align 4
+// CHECK19-NEXT:    [[TMP94:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.37, i32 0, i32 0), i64** [[TMP94]], align 4
+// CHECK19-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP95]], align 4
+// CHECK19-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP96]], align 4
+// CHECK19-NEXT:    [[TMP97:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS27]])
+// CHECK19-NEXT:    [[TMP98:%.*]] = icmp ne i32 [[TMP97]], 0
+// CHECK19-NEXT:    br i1 [[TMP98]], label [[OMP_OFFLOAD_FAILED28:%.*]], label [[OMP_OFFLOAD_CONT29:%.*]]
+// CHECK19:       omp_offload.failed28:
+// CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l142(i32 [[TMP76]], [10 x i32]* [[A]]) #[[ATTR3]]
+// CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT29]]
+// CHECK19:       omp_offload.cont29:
 // CHECK19-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/teams_distribute_private_codegen.cpp b/clang/test/OpenMP/teams_distribute_private_codegen.cpp
index 188f698..5173c98 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_private_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_private_codegen.cpp
@@ -242,9 +242,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -388,27 +405,44 @@ int main() {
 // CHECK1-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 8
 // CHECK1-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 8
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK1-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP4]]
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev
@@ -730,9 +764,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l93() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -874,27 +925,44 @@ int main() {
 // CHECK3-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 4
 // CHECK3-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK3-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP4]]
+// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev

diff --git a/clang/test/OpenMP/teams_distribute_reduction_codegen.cpp b/clang/test/OpenMP/teams_distribute_reduction_codegen.cpp
index c9389d3..ec5c0d0 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_reduction_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_reduction_codegen.cpp
@@ -107,9 +107,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l63.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l63.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l63(i64 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -272,9 +289,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP10:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[TMP10]], 0
-// CHECK1-NEXT:    br i1 [[TMP11]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP9]], i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i64 [[TMP2]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -439,9 +473,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l63.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l63.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK3-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l63(i32 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -602,9 +653,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP10:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[TMP10]], 0
-// CHECK3-NEXT:    br i1 [[TMP11]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP9]], i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK3-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i32 [[TMP2]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_simd_codegen.cpp b/clang/test/OpenMP/teams_distribute_simd_codegen.cpp
index 2fb2a1e..9f4822f 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_simd_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_simd_codegen.cpp
@@ -310,55 +310,89 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP39]], 1
 // CHECK1-NEXT:    [[TMP40:%.*]] = zext i32 [[ADD]] to i64
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP40]])
-// CHECK1-NEXT:    [[TMP41:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l30.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP36]], i32 1)
-// CHECK1-NEXT:    [[TMP42:%.*]] = icmp ne i32 [[TMP41]], 0
-// CHECK1-NEXT:    br i1 [[TMP42]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP34]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP35]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP36]], i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l30.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l30(i64 [[TMP2]], i64 [[TMP4]], [100 x i32]* @a, i64 [[TMP6]], i64 [[TMP8]]) #[[ATTR3:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
-// CHECK1-NEXT:    [[TMP43:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i32, i32* [[N_ADDR]], align 4
 // CHECK1-NEXT:    [[CONV8:%.*]] = bitcast i64* [[N_CASTED7]] to i32*
-// CHECK1-NEXT:    store i32 [[TMP43]], i32* [[CONV8]], align 4
-// CHECK1-NEXT:    [[TMP44:%.*]] = load i64, i64* [[N_CASTED7]], align 8
-// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP44]], i64* [[TMP46]], align 8
-// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i64*
-// CHECK1-NEXT:    store i64 [[TMP44]], i64* [[TMP48]], align 8
-// CHECK1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP49]], align 8
-// CHECK1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to [100 x i32]**
-// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP51]], align 8
-// CHECK1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 1
-// CHECK1-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to [100 x i32]**
-// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP53]], align 8
-// CHECK1-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 1
-// CHECK1-NEXT:    store i8* null, i8** [[TMP54]], align 8
-// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP57:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK1-NEXT:    store i32 [[TMP57]], i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK1-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
-// CHECK1-NEXT:    [[SUB15:%.*]] = sub nsw i32 [[TMP58]], 0
+// CHECK1-NEXT:    store i32 [[TMP51]], i32* [[CONV8]], align 4
+// CHECK1-NEXT:    [[TMP52:%.*]] = load i64, i64* [[N_CASTED7]], align 8
+// CHECK1-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP52]], i64* [[TMP54]], align 8
+// CHECK1-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i64*
+// CHECK1-NEXT:    store i64 [[TMP52]], i64* [[TMP56]], align 8
+// CHECK1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP57]], align 8
+// CHECK1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to [100 x i32]**
+// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP59]], align 8
+// CHECK1-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 1
+// CHECK1-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to [100 x i32]**
+// CHECK1-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP61]], align 8
+// CHECK1-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS11]], i64 0, i64 1
+// CHECK1-NEXT:    store i8* null, i8** [[TMP62]], align 8
+// CHECK1-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS9]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS10]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP65:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK1-NEXT:    store i32 [[TMP65]], i32* [[DOTCAPTURE_EXPR_13]], align 4
+// CHECK1-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_13]], align 4
+// CHECK1-NEXT:    [[SUB15:%.*]] = sub nsw i32 [[TMP66]], 0
 // CHECK1-NEXT:    [[DIV16:%.*]] = sdiv i32 [[SUB15]], 1
 // CHECK1-NEXT:    [[SUB17:%.*]] = sub nsw i32 [[DIV16]], 1
 // CHECK1-NEXT:    store i32 [[SUB17]], i32* [[DOTCAPTURE_EXPR_14]], align 4
-// CHECK1-NEXT:    [[TMP59:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_14]], align 4
-// CHECK1-NEXT:    [[ADD18:%.*]] = add nsw i32 [[TMP59]], 1
-// CHECK1-NEXT:    [[TMP60:%.*]] = zext i32 [[ADD18]] to i64
-// CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP60]])
-// CHECK1-NEXT:    [[TMP61:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP62:%.*]] = icmp ne i32 [[TMP61]], 0
-// CHECK1-NEXT:    br i1 [[TMP62]], label [[OMP_OFFLOAD_FAILED19:%.*]], label [[OMP_OFFLOAD_CONT20:%.*]]
-// CHECK1:       omp_offload.failed19:
-// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36(i64 [[TMP44]], [100 x i32]* @a) #[[ATTR3]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT20]]
-// CHECK1:       omp_offload.cont20:
-// CHECK1-NEXT:    [[TMP63:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i64 0, i64 0), align 4
-// CHECK1-NEXT:    ret i32 [[TMP63]]
+// CHECK1-NEXT:    [[TMP67:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_14]], align 4
+// CHECK1-NEXT:    [[ADD18:%.*]] = add nsw i32 [[TMP67]], 1
+// CHECK1-NEXT:    [[TMP68:%.*]] = zext i32 [[ADD18]] to i64
+// CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP68]])
+// CHECK1-NEXT:    [[KERNEL_ARGS19:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP69]], align 4
+// CHECK1-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 2, i32* [[TMP70]], align 4
+// CHECK1-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP63]], i8*** [[TMP71]], align 8
+// CHECK1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP64]], i8*** [[TMP72]], align 8
+// CHECK1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP73]], align 8
+// CHECK1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP74]], align 8
+// CHECK1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP75]], align 8
+// CHECK1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP76]], align 8
+// CHECK1-NEXT:    [[TMP77:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS19]])
+// CHECK1-NEXT:    [[TMP78:%.*]] = icmp ne i32 [[TMP77]], 0
+// CHECK1-NEXT:    br i1 [[TMP78]], label [[OMP_OFFLOAD_FAILED20:%.*]], label [[OMP_OFFLOAD_CONT21:%.*]]
+// CHECK1:       omp_offload.failed20:
+// CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36(i64 [[TMP52]], [100 x i32]* @a) #[[ATTR3]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT21]]
+// CHECK1:       omp_offload.cont21:
+// CHECK1-NEXT:    [[TMP79:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i64 0, i64 0), align 4
+// CHECK1-NEXT:    ret i32 [[TMP79]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l30
@@ -715,54 +749,88 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP39]], 1
 // CHECK3-NEXT:    [[TMP40:%.*]] = zext i32 [[ADD]] to i64
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP40]])
-// CHECK3-NEXT:    [[TMP41:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l30.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 [[TMP36]], i32 1)
-// CHECK3-NEXT:    [[TMP42:%.*]] = icmp ne i32 [[TMP41]], 0
-// CHECK3-NEXT:    br i1 [[TMP42]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP34]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP35]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP36]], i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l30.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l30(i32 [[TMP2]], i32 [[TMP4]], [100 x i32]* @a, i32 [[TMP6]], i32 [[TMP8]]) #[[ATTR3:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
-// CHECK3-NEXT:    [[TMP43:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    store i32 [[TMP43]], i32* [[N_CASTED4]], align 4
-// CHECK3-NEXT:    [[TMP44:%.*]] = load i32, i32* [[N_CASTED4]], align 4
-// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP44]], i32* [[TMP46]], align 4
-// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
-// CHECK3-NEXT:    store i32 [[TMP44]], i32* [[TMP48]], align 4
-// CHECK3-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP49]], align 4
-// CHECK3-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to [100 x i32]**
-// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP51]], align 4
-// CHECK3-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK3-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to [100 x i32]**
-// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP53]], align 4
-// CHECK3-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 1
-// CHECK3-NEXT:    store i8* null, i8** [[TMP54]], align 4
-// CHECK3-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP57:%.*]] = load i32, i32* [[N_ADDR]], align 4
-// CHECK3-NEXT:    store i32 [[TMP57]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK3-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK3-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP58]], 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    store i32 [[TMP51]], i32* [[N_CASTED4]], align 4
+// CHECK3-NEXT:    [[TMP52:%.*]] = load i32, i32* [[N_CASTED4]], align 4
+// CHECK3-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP54:%.*]] = bitcast i8** [[TMP53]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP52]], i32* [[TMP54]], align 4
+// CHECK3-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32*
+// CHECK3-NEXT:    store i32 [[TMP52]], i32* [[TMP56]], align 4
+// CHECK3-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP57]], align 4
+// CHECK3-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP59:%.*]] = bitcast i8** [[TMP58]] to [100 x i32]**
+// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP59]], align 4
+// CHECK3-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK3-NEXT:    [[TMP61:%.*]] = bitcast i8** [[TMP60]] to [100 x i32]**
+// CHECK3-NEXT:    store [100 x i32]* @a, [100 x i32]** [[TMP61]], align 4
+// CHECK3-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i8* null, i8** [[TMP62]], align 4
+// CHECK3-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP65:%.*]] = load i32, i32* [[N_ADDR]], align 4
+// CHECK3-NEXT:    store i32 [[TMP65]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK3-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK3-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP66]], 0
 // CHECK3-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK3-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK3-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK3-NEXT:    [[TMP59:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK3-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP59]], 1
-// CHECK3-NEXT:    [[TMP60:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP60]])
-// CHECK3-NEXT:    [[TMP61:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36.region_id, i32 2, i8** [[TMP55]], i8** [[TMP56]], i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP62:%.*]] = icmp ne i32 [[TMP61]], 0
-// CHECK3-NEXT:    br i1 [[TMP62]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK3:       omp_offload.failed15:
-// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36(i32 [[TMP44]], [100 x i32]* @a) #[[ATTR3]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK3:       omp_offload.cont16:
-// CHECK3-NEXT:    [[TMP63:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i32 0, i32 0), align 4
-// CHECK3-NEXT:    ret i32 [[TMP63]]
+// CHECK3-NEXT:    [[TMP67:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK3-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP67]], 1
+// CHECK3-NEXT:    [[TMP68:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP68]])
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP69]], align 4
+// CHECK3-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 2, i32* [[TMP70]], align 4
+// CHECK3-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP63]], i8*** [[TMP71]], align 4
+// CHECK3-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP64]], i8*** [[TMP72]], align 4
+// CHECK3-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK3-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([2 x i64], [2 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP74]], align 4
+// CHECK3-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK3-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP76]], align 4
+// CHECK3-NEXT:    [[TMP77:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP78:%.*]] = icmp ne i32 [[TMP77]], 0
+// CHECK3-NEXT:    br i1 [[TMP78]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
+// CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l36(i32 [[TMP52]], [100 x i32]* @a) #[[ATTR3]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[TMP79:%.*]] = load i32, i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i32 0, i32 0), align 4
+// CHECK3-NEXT:    ret i32 [[TMP79]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z21teams_argument_globali_l30
@@ -1353,18 +1421,35 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK9-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l75.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l75.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK9-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l75(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 0
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP33]])
-// CHECK9-NEXT:    ret i32 [[TMP32]]
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP41]])
+// CHECK9-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l75
@@ -1564,18 +1649,35 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK11-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK11-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l75.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK11-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l75.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK11-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l75(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i32 0
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP33]])
-// CHECK11-NEXT:    ret i32 [[TMP32]]
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP41]])
+// CHECK11-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z15teams_local_argv_l75
@@ -1918,17 +2020,34 @@ int main (int argc, char **argv) {
 // CHECK17-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK17-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK17-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* [[TMP26]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK17-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK17-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK17-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 3, i32* [[TMP28]], align 4
+// CHECK17-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP24]], i8*** [[TMP29]], align 8
+// CHECK17-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP25]], i8*** [[TMP30]], align 8
+// CHECK17-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP26]], i64** [[TMP31]], align 8
+// CHECK17-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP32]], align 8
+// CHECK17-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK17-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP34]], align 8
+// CHECK17-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK17-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
 // CHECK17-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK17-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A2]], i64 0, i64 0
-// CHECK17-NEXT:    [[TMP29:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK17-NEXT:    ret i32 [[TMP29]]
+// CHECK17-NEXT:    [[TMP37:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK17-NEXT:    ret i32 [[TMP37]]
 //
 //
 // CHECK17-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123
@@ -2086,17 +2205,34 @@ int main (int argc, char **argv) {
 // CHECK19-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK19-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK19-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* [[TMP26]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK19-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK19-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK19-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 3, i32* [[TMP28]], align 4
+// CHECK19-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP24]], i8*** [[TMP29]], align 4
+// CHECK19-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP25]], i8*** [[TMP30]], align 4
+// CHECK19-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP26]], i64** [[TMP31]], align 4
+// CHECK19-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP32]], align 4
+// CHECK19-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK19-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP34]], align 4
+// CHECK19-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK19-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
 // CHECK19-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK19-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A2]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP29:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK19-NEXT:    ret i32 [[TMP29]]
+// CHECK19-NEXT:    [[TMP37:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK19-NEXT:    ret i32 [[TMP37]]
 //
 //
 // CHECK19-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123
@@ -2253,17 +2389,34 @@ int main (int argc, char **argv) {
 // CHECK21-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK21-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK21-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK21-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* [[TMP26]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK21-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK21-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK21-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK21-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK21-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK21-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK21-NEXT:    store i32 3, i32* [[TMP28]], align 4
+// CHECK21-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK21-NEXT:    store i8** [[TMP24]], i8*** [[TMP29]], align 8
+// CHECK21-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK21-NEXT:    store i8** [[TMP25]], i8*** [[TMP30]], align 8
+// CHECK21-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK21-NEXT:    store i64* [[TMP26]], i64** [[TMP31]], align 8
+// CHECK21-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK21-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP32]], align 8
+// CHECK21-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK21-NEXT:    store i8** null, i8*** [[TMP33]], align 8
+// CHECK21-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK21-NEXT:    store i8** null, i8*** [[TMP34]], align 8
+// CHECK21-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK21-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK21-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK21:       omp_offload.failed:
 // CHECK21-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK21-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK21:       omp_offload.cont:
 // CHECK21-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK21-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A2]], i64 0, i64 0
-// CHECK21-NEXT:    [[TMP29:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK21-NEXT:    ret i32 [[TMP29]]
+// CHECK21-NEXT:    [[TMP37:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK21-NEXT:    ret i32 [[TMP37]]
 //
 //
 // CHECK21-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123
@@ -2473,17 +2626,34 @@ int main (int argc, char **argv) {
 // CHECK23-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK23-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
 // CHECK23-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK23-NEXT:    [[TMP27:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123.region_id, i32 3, i8** [[TMP24]], i8** [[TMP25]], i64* [[TMP26]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK23-NEXT:    [[TMP28:%.*]] = icmp ne i32 [[TMP27]], 0
-// CHECK23-NEXT:    br i1 [[TMP28]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK23-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK23-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK23-NEXT:    store i32 1, i32* [[TMP27]], align 4
+// CHECK23-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK23-NEXT:    store i32 3, i32* [[TMP28]], align 4
+// CHECK23-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK23-NEXT:    store i8** [[TMP24]], i8*** [[TMP29]], align 4
+// CHECK23-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK23-NEXT:    store i8** [[TMP25]], i8*** [[TMP30]], align 4
+// CHECK23-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK23-NEXT:    store i64* [[TMP26]], i64** [[TMP31]], align 4
+// CHECK23-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK23-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP32]], align 4
+// CHECK23-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK23-NEXT:    store i8** null, i8*** [[TMP33]], align 4
+// CHECK23-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK23-NEXT:    store i8** null, i8*** [[TMP34]], align 4
+// CHECK23-NEXT:    [[TMP35:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK23-NEXT:    [[TMP36:%.*]] = icmp ne i32 [[TMP35]], 0
+// CHECK23-NEXT:    br i1 [[TMP36]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK23:       omp_offload.failed:
 // CHECK23-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK23-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK23:       omp_offload.cont:
 // CHECK23-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK23-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A2]], i32 0, i32 0
-// CHECK23-NEXT:    [[TMP29:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK23-NEXT:    ret i32 [[TMP29]]
+// CHECK23-NEXT:    [[TMP37:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK23-NEXT:    ret i32 [[TMP37]]
 //
 //
 // CHECK23-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l123
@@ -3018,20 +3188,37 @@ int main (int argc, char **argv) {
 // CHECK33-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK33-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK33-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK33-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK33-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK33-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK33-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK33-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK33-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK33-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK33-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK33-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK33-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK33-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK33-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK33-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK33-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK33-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK33-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK33-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK33-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK33-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK33-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK33-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK33-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK33-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK33:       omp_offload.failed:
 // CHECK33-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK33-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK33:       omp_offload.cont:
-// CHECK33-NEXT:    [[TMP32:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK33-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP32]])
+// CHECK33-NEXT:    [[TMP40:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK33-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP40]])
 // CHECK33-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK33-NEXT:    [[TMP33:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK33-NEXT:    call void @llvm.stackrestore(i8* [[TMP33]])
-// CHECK33-NEXT:    [[TMP34:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK33-NEXT:    ret i32 [[TMP34]]
+// CHECK33-NEXT:    [[TMP41:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK33-NEXT:    call void @llvm.stackrestore(i8* [[TMP41]])
+// CHECK33-NEXT:    [[TMP42:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK33-NEXT:    ret i32 [[TMP42]]
 //
 //
 // CHECK33-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192
@@ -3209,9 +3396,26 @@ int main (int argc, char **argv) {
 // CHECK33-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK33-NEXT:    [[TMP21:%.*]] = load i32, i32* [[TE]], align 4
 // CHECK33-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK33-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l181.region_id, i32 3, i8** [[TMP19]], i8** [[TMP20]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 [[TMP21]], i32 1)
-// CHECK33-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK33-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK33-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK33-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK33-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK33-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK33-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK33-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK33-NEXT:    store i8** [[TMP19]], i8*** [[TMP24]], align 8
+// CHECK33-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK33-NEXT:    store i8** [[TMP20]], i8*** [[TMP25]], align 8
+// CHECK33-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK33-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK33-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK33-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK33-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK33-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK33-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK33-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK33-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP21]], i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l181.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK33-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK33-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK33:       omp_offload.failed:
 // CHECK33-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l181(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK33-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3393,20 +3597,37 @@ int main (int argc, char **argv) {
 // CHECK35-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK35-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK35-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK35-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK35-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK35-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK35-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK35-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK35-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK35-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK35-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK35-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK35-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK35-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK35-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK35-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK35-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK35-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK35-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK35-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK35-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK35-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK35-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK35-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK35-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK35-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK35:       omp_offload.failed:
 // CHECK35-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK35-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK35:       omp_offload.cont:
-// CHECK35-NEXT:    [[TMP32:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK35-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP32]])
+// CHECK35-NEXT:    [[TMP40:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK35-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP40]])
 // CHECK35-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK35-NEXT:    [[TMP33:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK35-NEXT:    call void @llvm.stackrestore(i8* [[TMP33]])
-// CHECK35-NEXT:    [[TMP34:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK35-NEXT:    ret i32 [[TMP34]]
+// CHECK35-NEXT:    [[TMP41:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK35-NEXT:    call void @llvm.stackrestore(i8* [[TMP41]])
+// CHECK35-NEXT:    [[TMP42:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK35-NEXT:    ret i32 [[TMP42]]
 //
 //
 // CHECK35-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192
@@ -3580,9 +3801,26 @@ int main (int argc, char **argv) {
 // CHECK35-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK35-NEXT:    [[TMP21:%.*]] = load i32, i32* [[TE]], align 4
 // CHECK35-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK35-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l181.region_id, i32 3, i8** [[TMP19]], i8** [[TMP20]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 [[TMP21]], i32 1)
-// CHECK35-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK35-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK35-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK35-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK35-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK35-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK35-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK35-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK35-NEXT:    store i8** [[TMP19]], i8*** [[TMP24]], align 4
+// CHECK35-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK35-NEXT:    store i8** [[TMP20]], i8*** [[TMP25]], align 4
+// CHECK35-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK35-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK35-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK35-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK35-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK35-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK35-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK35-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK35-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP21]], i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l181.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK35-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK35-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK35:       omp_offload.failed:
 // CHECK35-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l181(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK35-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -3775,20 +4013,37 @@ int main (int argc, char **argv) {
 // CHECK37-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP35]], 1
 // CHECK37-NEXT:    [[TMP36:%.*]] = zext i32 [[ADD]] to i64
 // CHECK37-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP36]])
-// CHECK37-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192.region_id, i32 4, i8** [[TMP30]], i8** [[TMP31]], i64* [[TMP32]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK37-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
-// CHECK37-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK37-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK37-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK37-NEXT:    store i32 1, i32* [[TMP37]], align 4
+// CHECK37-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK37-NEXT:    store i32 4, i32* [[TMP38]], align 4
+// CHECK37-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK37-NEXT:    store i8** [[TMP30]], i8*** [[TMP39]], align 8
+// CHECK37-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK37-NEXT:    store i8** [[TMP31]], i8*** [[TMP40]], align 8
+// CHECK37-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK37-NEXT:    store i64* [[TMP32]], i64** [[TMP41]], align 8
+// CHECK37-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK37-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP42]], align 8
+// CHECK37-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK37-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK37-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK37-NEXT:    store i8** null, i8*** [[TMP44]], align 8
+// CHECK37-NEXT:    [[TMP45:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK37-NEXT:    [[TMP46:%.*]] = icmp ne i32 [[TMP45]], 0
+// CHECK37-NEXT:    br i1 [[TMP46]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK37:       omp_offload.failed:
 // CHECK37-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192(i64 [[TMP4]], i64 [[TMP6]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK37-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK37:       omp_offload.cont:
-// CHECK37-NEXT:    [[TMP39:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK37-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP39]])
+// CHECK37-NEXT:    [[TMP47:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK37-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP47]])
 // CHECK37-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK37-NEXT:    [[TMP40:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK37-NEXT:    call void @llvm.stackrestore(i8* [[TMP40]])
-// CHECK37-NEXT:    [[TMP41:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK37-NEXT:    ret i32 [[TMP41]]
+// CHECK37-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK37-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
+// CHECK37-NEXT:    [[TMP49:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK37-NEXT:    ret i32 [[TMP49]]
 //
 //
 // CHECK37-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192
@@ -4016,9 +4271,26 @@ int main (int argc, char **argv) {
 // CHECK37-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK37-NEXT:    [[TMP21:%.*]] = load i32, i32* [[TE]], align 4
 // CHECK37-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK37-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l181.region_id, i32 3, i8** [[TMP19]], i8** [[TMP20]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 [[TMP21]], i32 1)
-// CHECK37-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK37-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK37-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK37-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK37-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK37-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK37-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK37-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK37-NEXT:    store i8** [[TMP19]], i8*** [[TMP24]], align 8
+// CHECK37-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK37-NEXT:    store i8** [[TMP20]], i8*** [[TMP25]], align 8
+// CHECK37-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK37-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP26]], align 8
+// CHECK37-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK37-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK37-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK37-NEXT:    store i8** null, i8*** [[TMP28]], align 8
+// CHECK37-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK37-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK37-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP21]], i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l181.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK37-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK37-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK37:       omp_offload.failed:
 // CHECK37-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l181(i64 [[TMP1]], i64 [[TMP3]], [10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK37-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -4212,20 +4484,37 @@ int main (int argc, char **argv) {
 // CHECK39-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP35]], 1
 // CHECK39-NEXT:    [[TMP36:%.*]] = zext i32 [[ADD]] to i64
 // CHECK39-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP36]])
-// CHECK39-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192.region_id, i32 4, i8** [[TMP30]], i8** [[TMP31]], i64* [[TMP32]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK39-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
-// CHECK39-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK39-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK39-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK39-NEXT:    store i32 1, i32* [[TMP37]], align 4
+// CHECK39-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK39-NEXT:    store i32 4, i32* [[TMP38]], align 4
+// CHECK39-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK39-NEXT:    store i8** [[TMP30]], i8*** [[TMP39]], align 4
+// CHECK39-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK39-NEXT:    store i8** [[TMP31]], i8*** [[TMP40]], align 4
+// CHECK39-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK39-NEXT:    store i64* [[TMP32]], i64** [[TMP41]], align 4
+// CHECK39-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK39-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP42]], align 4
+// CHECK39-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK39-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK39-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK39-NEXT:    store i8** null, i8*** [[TMP44]], align 4
+// CHECK39-NEXT:    [[TMP45:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK39-NEXT:    [[TMP46:%.*]] = icmp ne i32 [[TMP45]], 0
+// CHECK39-NEXT:    br i1 [[TMP46]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK39:       omp_offload.failed:
 // CHECK39-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192(i32 [[TMP3]], i32 [[TMP5]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK39-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK39:       omp_offload.cont:
-// CHECK39-NEXT:    [[TMP39:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK39-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP39]])
+// CHECK39-NEXT:    [[TMP47:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK39-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP47]])
 // CHECK39-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK39-NEXT:    [[TMP40:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK39-NEXT:    call void @llvm.stackrestore(i8* [[TMP40]])
-// CHECK39-NEXT:    [[TMP41:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK39-NEXT:    ret i32 [[TMP41]]
+// CHECK39-NEXT:    [[TMP48:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK39-NEXT:    call void @llvm.stackrestore(i8* [[TMP48]])
+// CHECK39-NEXT:    [[TMP49:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK39-NEXT:    ret i32 [[TMP49]]
 //
 //
 // CHECK39-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l192
@@ -4447,9 +4736,26 @@ int main (int argc, char **argv) {
 // CHECK39-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK39-NEXT:    [[TMP21:%.*]] = load i32, i32* [[TE]], align 4
 // CHECK39-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK39-NEXT:    [[TMP22:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l181.region_id, i32 3, i8** [[TMP19]], i8** [[TMP20]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 [[TMP21]], i32 1)
-// CHECK39-NEXT:    [[TMP23:%.*]] = icmp ne i32 [[TMP22]], 0
-// CHECK39-NEXT:    br i1 [[TMP23]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK39-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK39-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK39-NEXT:    store i32 1, i32* [[TMP22]], align 4
+// CHECK39-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK39-NEXT:    store i32 3, i32* [[TMP23]], align 4
+// CHECK39-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK39-NEXT:    store i8** [[TMP19]], i8*** [[TMP24]], align 4
+// CHECK39-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK39-NEXT:    store i8** [[TMP20]], i8*** [[TMP25]], align 4
+// CHECK39-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK39-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP26]], align 4
+// CHECK39-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK39-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK39-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK39-NEXT:    store i8** null, i8*** [[TMP28]], align 4
+// CHECK39-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK39-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK39-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 [[TMP21]], i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l181.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK39-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
+// CHECK39-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK39:       omp_offload.failed:
 // CHECK39-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l181(i32 [[TMP1]], i32 [[TMP3]], [10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK39-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_simd_collapse_codegen.cpp b/clang/test/OpenMP/teams_distribute_simd_collapse_codegen.cpp
index 1754486..abc8c22 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_simd_collapse_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_simd_collapse_codegen.cpp
@@ -129,9 +129,26 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 56088)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -139,8 +156,8 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i64 0, i64 0
 // CHECK1-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP9]]
+// CHECK1-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -281,9 +298,26 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 56088)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -291,8 +325,8 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[A3:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x [456 x i32]], [123 x [456 x i32]]* [[A3]], i32 0, i32 0
 // CHECK3-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds [456 x i32], [456 x i32]* [[ARRAYIDX]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP9]]
+// CHECK3-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ARRAYIDX4]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP17]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -652,20 +686,37 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP46:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_4]], align 8
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP46]], 1
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK9-NEXT:    [[TMP47:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83.region_id, i32 5, i8** [[TMP39]], i8** [[TMP40]], i64* [[TMP41]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP48:%.*]] = icmp ne i32 [[TMP47]], 0
-// CHECK9-NEXT:    br i1 [[TMP48]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP47]], align 4
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP48]], align 4
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 8
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP40]], i8*** [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP41]], i64** [[TMP51]], align 8
+// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP56:%.*]] = icmp ne i32 [[TMP55]], 0
+// CHECK9-NEXT:    br i1 [[TMP56]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83(i64 [[TMP7]], i64 [[TMP9]], i64 [[TMP1]], i64 [[TMP3]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP49:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP49]])
+// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef signext [[TMP57]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP50:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP50]])
-// CHECK9-NEXT:    [[TMP51:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP51]]
+// CHECK9-NEXT:    [[TMP58:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP58]])
+// CHECK9-NEXT:    [[TMP59:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP59]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83
@@ -878,9 +929,26 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 20)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l69.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l69.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l69([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1103,20 +1171,37 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP45:%.*]] = load i64, i64* [[DOTCAPTURE_EXPR_3]], align 8
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i64 [[TMP45]], 1
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[ADD]])
-// CHECK11-NEXT:    [[TMP46:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83.region_id, i32 5, i8** [[TMP38]], i8** [[TMP39]], i64* [[TMP40]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP47:%.*]] = icmp ne i32 [[TMP46]], 0
-// CHECK11-NEXT:    br i1 [[TMP47]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP38]], i8*** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP39]], i8*** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP40]], i64** [[TMP50]], align 4
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP55:%.*]] = icmp ne i32 [[TMP54]], 0
+// CHECK11-NEXT:    br i1 [[TMP55]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83(i32 [[TMP5]], i32 [[TMP7]], i32 [[TMP0]], i32 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP48:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP48]])
+// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10ELi2EEiT_(i32 noundef [[TMP56]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP49]])
-// CHECK11-NEXT:    [[TMP50:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP50]]
+// CHECK11-NEXT:    [[TMP57:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP57]])
+// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP58]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l83
@@ -1325,9 +1410,26 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 20)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l69.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l69.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10ELi2EEiT__l69([10 x [2 x i32]]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_distribute_simd_dist_schedule_codegen.cpp b/clang/test/OpenMP/teams_distribute_simd_dist_schedule_codegen.cpp
index 4512d93..08249fe 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_simd_dist_schedule_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_simd_dist_schedule_codegen.cpp
@@ -151,10 +151,10 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 8
 // CHECK1-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 8
-// CHECK1-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK1-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 8
+// CHECK1-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 8
 // CHECK1-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -169,55 +169,106 @@ int main (int argc, char **argv) {
 // CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK1-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK1-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 8
-// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 8
-// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 8
+// CHECK1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 8
+// CHECK1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK1:       omp_offload.failed7:
+// CHECK1-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK1-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK1-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK1:       omp_offload.failed8:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK1:       omp_offload.cont8:
-// CHECK1-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 8
-// CHECK1-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK1-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 8
-// CHECK1-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i64 0, i64 0
-// CHECK1-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK1:       omp_offload.cont9:
+// CHECK1-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK1-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK1-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i64 0, i64 0
+// CHECK1-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK1-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK1-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK1-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK1:       omp_offload.failed14:
+// CHECK1-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK1-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK1-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK1-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK1-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK1-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK1-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK1-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK1-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK1-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK1-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK1:       omp_offload.failed16:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK1:       omp_offload.cont15:
-// CHECK1-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i64 0, i64 0
-// CHECK1-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP27]]
+// CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK1:       omp_offload.cont17:
+// CHECK1-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i64 0, i64 0
+// CHECK1-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -519,10 +570,10 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[DOTOFFLOAD_PTRS4:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS5:%.*]] = alloca [1 x i8*], align 4
 // CHECK3-NEXT:    [[_TMP6:%.*]] = alloca i32, align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS10:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS11:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS12:%.*]] = alloca [1 x i8*], align 4
-// CHECK3-NEXT:    [[_TMP13:%.*]] = alloca i32, align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_BASEPTRS11:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_PTRS12:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[DOTOFFLOAD_MAPPERS13:%.*]] = alloca [1 x i8*], align 4
+// CHECK3-NEXT:    [[_TMP14:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store %struct.SS* [[THIS]], %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[THIS1:%.*]] = load %struct.SS*, %struct.SS** [[THIS_ADDR]], align 4
 // CHECK3-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_SS:%.*]], %struct.SS* [[THIS1]], i32 0, i32 0
@@ -537,55 +588,106 @@ int main (int argc, char **argv) {
 // CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK3-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK3-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28(%struct.SS* [[THIS1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP10]], align 4
-// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP12]], align 4
-// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP18]], align 4
+// CHECK3-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A2]], [123 x i32]** [[TMP20]], align 4
+// CHECK3-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS5]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS3]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS4]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK3:       omp_offload.failed7:
+// CHECK3-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK3-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK3-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK3-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK3-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK3:       omp_offload.failed8:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l33(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK3:       omp_offload.cont8:
-// CHECK3-NEXT:    [[A9:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to %struct.SS**
-// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP19]], align 4
-// CHECK3-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [123 x i32]**
-// CHECK3-NEXT:    store [123 x i32]* [[A9]], [123 x i32]** [[TMP21]], align 4
-// CHECK3-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS12]], i32 0, i32 0
-// CHECK3-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK3-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS10]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK3:       omp_offload.cont9:
+// CHECK3-NEXT:    [[A10:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to %struct.SS**
+// CHECK3-NEXT:    store %struct.SS* [[THIS1]], %struct.SS** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [123 x i32]**
+// CHECK3-NEXT:    store [123 x i32]* [[A10]], [123 x i32]** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS13]], i32 0, i32 0
+// CHECK3-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS11]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS12]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 123)
-// CHECK3-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK3-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED14:%.*]], label [[OMP_OFFLOAD_CONT15:%.*]]
-// CHECK3:       omp_offload.failed14:
+// CHECK3-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK3-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK3-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK3-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK3-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK3-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK3-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK3-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK3-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK3-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK3-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK3-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK3:       omp_offload.failed16:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l38(%struct.SS* [[THIS1]]) #[[ATTR2]]
-// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT15]]
-// CHECK3:       omp_offload.cont15:
-// CHECK3-NEXT:    [[A16:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
-// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A16]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP27:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP27]]
+// CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK3:       omp_offload.cont17:
+// CHECK3-NEXT:    [[A18:%.*]] = getelementptr inbounds [[STRUCT_SS]], %struct.SS* [[THIS1]], i32 0, i32 0
+// CHECK3-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [123 x i32], [123 x i32]* [[A18]], i32 0, i32 0
+// CHECK3-NEXT:    [[TMP51:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSIiLi123ELx456EE3fooEv_l28
@@ -1136,14 +1238,14 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[_TMP9:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_11:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[N_CASTED18:%.*]] = alloca i64, align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS20:%.*]] = alloca [3 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS21:%.*]] = alloca [3 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS22:%.*]] = alloca [3 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES23:%.*]] = alloca [3 x i64], align 8
-// CHECK9-NEXT:    [[_TMP24:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[N_CASTED19:%.*]] = alloca i64, align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS21:%.*]] = alloca [3 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS22:%.*]] = alloca [3 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS23:%.*]] = alloca [3 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_SIZES24:%.*]] = alloca [3 x i64], align 8
+// CHECK9-NEXT:    [[_TMP25:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    [[DOTCAPTURE_EXPR_26:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTCAPTURE_EXPR_27:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 8
@@ -1201,128 +1303,179 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK9-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK9-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK9-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK9-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100(i64 [[TMP4]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
 // CHECK9-NEXT:    [[CONV4:%.*]] = bitcast i64* [[N_CASTED3]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP32]], i32* [[CONV4]], align 4
-// CHECK9-NEXT:    [[TMP33:%.*]] = load i64, i64* [[N_CASTED3]], align 8
-// CHECK9-NEXT:    [[TMP34:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP35:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP35]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i64 24, i1 false)
-// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP37]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP33]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP42:%.*]] = bitcast i8** [[TMP41]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP42]], align 8
-// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP44:%.*]] = bitcast i8** [[TMP43]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP44]], align 8
-// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP45]], align 8
-// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP47]], align 8
-// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP49:%.*]] = bitcast i8** [[TMP48]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP49]], align 8
-// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP34]], i64* [[TMP50]], align 8
-// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP51]], align 8
-// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP55:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP55]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[TMP56:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP56]], 0
+// CHECK9-NEXT:    store i32 [[TMP40]], i32* [[CONV4]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = load i64, i64* [[N_CASTED3]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP43:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES8]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP43]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i64 24, i1 false)
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP41]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP50]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP52:%.*]] = bitcast i8** [[TMP51]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP52]], align 8
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP55:%.*]] = bitcast i8** [[TMP54]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP57:%.*]] = bitcast i8** [[TMP56]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP42]], i64* [[TMP58]], align 8
+// CHECK9-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS7]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP59]], align 8
+// CHECK9-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS6]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK9-NEXT:    [[SUB12:%.*]] = sub nsw i32 [[TMP64]], 0
 // CHECK9-NEXT:    [[DIV13:%.*]] = sdiv i32 [[SUB12]], 1
 // CHECK9-NEXT:    [[SUB14:%.*]] = sub nsw i32 [[DIV13]], 1
 // CHECK9-NEXT:    store i32 [[SUB14]], i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
-// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP57]], 1
-// CHECK9-NEXT:    [[TMP58:%.*]] = zext i32 [[ADD15]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP58]])
-// CHECK9-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105.region_id, i32 3, i8** [[TMP52]], i8** [[TMP53]], i64* [[TMP54]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
-// CHECK9-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
-// CHECK9:       omp_offload.failed16:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105(i64 [[TMP33]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
-// CHECK9:       omp_offload.cont17:
-// CHECK9-NEXT:    [[TMP61:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    [[CONV19:%.*]] = bitcast i64* [[N_CASTED18]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP61]], i32* [[CONV19]], align 4
-// CHECK9-NEXT:    [[TMP62:%.*]] = load i64, i64* [[N_CASTED18]], align 8
-// CHECK9-NEXT:    [[TMP63:%.*]] = mul nuw i64 [[TMP1]], 4
-// CHECK9-NEXT:    [[TMP64:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES23]] to i8*
-// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP64]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.5 to i8*), i64 24, i1 false)
-// CHECK9-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP66:%.*]] = bitcast i8** [[TMP65]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP62]], i64* [[TMP66]], align 8
-// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP62]], i64* [[TMP68]], align 8
-// CHECK9-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP69]], align 8
-// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP71:%.*]] = bitcast i8** [[TMP70]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP71]], align 8
-// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 1
-// CHECK9-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP73]], align 8
-// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 1
-// CHECK9-NEXT:    store i8* null, i8** [[TMP74]], align 8
-// CHECK9-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP76:%.*]] = bitcast i8** [[TMP75]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP76]], align 8
-// CHECK9-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 2
-// CHECK9-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32**
-// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP78]], align 8
-// CHECK9-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 2
-// CHECK9-NEXT:    store i64 [[TMP63]], i64* [[TMP79]], align 8
-// CHECK9-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS22]], i64 0, i64 2
-// CHECK9-NEXT:    store i8* null, i8** [[TMP80]], align 8
-// CHECK9-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS20]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS21]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES23]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP84:%.*]] = load i32, i32* [[N]], align 4
-// CHECK9-NEXT:    store i32 [[TMP84]], i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK9-NEXT:    [[TMP85:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
-// CHECK9-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[TMP85]], 0
-// CHECK9-NEXT:    [[DIV28:%.*]] = sdiv i32 [[SUB27]], 1
-// CHECK9-NEXT:    [[SUB29:%.*]] = sub nsw i32 [[DIV28]], 1
-// CHECK9-NEXT:    store i32 [[SUB29]], i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK9-NEXT:    [[TMP86:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
-// CHECK9-NEXT:    [[ADD30:%.*]] = add nsw i32 [[TMP86]], 1
-// CHECK9-NEXT:    [[TMP87:%.*]] = zext i32 [[ADD30]] to i64
-// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP87]])
-// CHECK9-NEXT:    [[TMP88:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110.region_id, i32 3, i8** [[TMP81]], i8** [[TMP82]], i64* [[TMP83]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP89:%.*]] = icmp ne i32 [[TMP88]], 0
-// CHECK9-NEXT:    br i1 [[TMP89]], label [[OMP_OFFLOAD_FAILED31:%.*]], label [[OMP_OFFLOAD_CONT32:%.*]]
-// CHECK9:       omp_offload.failed31:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110(i64 [[TMP62]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT32]]
-// CHECK9:       omp_offload.cont32:
-// CHECK9-NEXT:    [[TMP90:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP90]])
+// CHECK9-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_11]], align 4
+// CHECK9-NEXT:    [[ADD15:%.*]] = add nsw i32 [[TMP65]], 1
+// CHECK9-NEXT:    [[TMP66:%.*]] = zext i32 [[ADD15]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP66]])
+// CHECK9-NEXT:    [[KERNEL_ARGS16:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP67]], align 4
+// CHECK9-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP68]], align 4
+// CHECK9-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP60]], i8*** [[TMP69]], align 8
+// CHECK9-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 8
+// CHECK9-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP62]], i64** [[TMP71]], align 8
+// CHECK9-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP72]], align 8
+// CHECK9-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP73]], align 8
+// CHECK9-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP74]], align 8
+// CHECK9-NEXT:    [[TMP75:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS16]])
+// CHECK9-NEXT:    [[TMP76:%.*]] = icmp ne i32 [[TMP75]], 0
+// CHECK9-NEXT:    br i1 [[TMP76]], label [[OMP_OFFLOAD_FAILED17:%.*]], label [[OMP_OFFLOAD_CONT18:%.*]]
+// CHECK9:       omp_offload.failed17:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105(i64 [[TMP41]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT18]]
+// CHECK9:       omp_offload.cont18:
+// CHECK9-NEXT:    [[TMP77:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    [[CONV20:%.*]] = bitcast i64* [[N_CASTED19]] to i32*
+// CHECK9-NEXT:    store i32 [[TMP77]], i32* [[CONV20]], align 4
+// CHECK9-NEXT:    [[TMP78:%.*]] = load i64, i64* [[N_CASTED19]], align 8
+// CHECK9-NEXT:    [[TMP79:%.*]] = mul nuw i64 [[TMP1]], 4
+// CHECK9-NEXT:    [[TMP80:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES24]] to i8*
+// CHECK9-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP80]], i8* align 8 bitcast ([3 x i64]* @.offload_sizes.5 to i8*), i64 24, i1 false)
+// CHECK9-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP82:%.*]] = bitcast i8** [[TMP81]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP78]], i64* [[TMP82]], align 8
+// CHECK9-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP78]], i64* [[TMP84]], align 8
+// CHECK9-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP85]], align 8
+// CHECK9-NEXT:    [[TMP86:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP87:%.*]] = bitcast i8** [[TMP86]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP87]], align 8
+// CHECK9-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 1
+// CHECK9-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP1]], i64* [[TMP89]], align 8
+// CHECK9-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 1
+// CHECK9-NEXT:    store i8* null, i8** [[TMP90]], align 8
+// CHECK9-NEXT:    [[TMP91:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP92:%.*]] = bitcast i8** [[TMP91]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP92]], align 8
+// CHECK9-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 2
+// CHECK9-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32**
+// CHECK9-NEXT:    store i32* [[VLA]], i32** [[TMP94]], align 8
+// CHECK9-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 2
+// CHECK9-NEXT:    store i64 [[TMP79]], i64* [[TMP95]], align 8
+// CHECK9-NEXT:    [[TMP96:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS23]], i64 0, i64 2
+// CHECK9-NEXT:    store i8* null, i8** [[TMP96]], align 8
+// CHECK9-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS21]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS22]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES24]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP100:%.*]] = load i32, i32* [[N]], align 4
+// CHECK9-NEXT:    store i32 [[TMP100]], i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK9-NEXT:    [[TMP101:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_26]], align 4
+// CHECK9-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[TMP101]], 0
+// CHECK9-NEXT:    [[DIV29:%.*]] = sdiv i32 [[SUB28]], 1
+// CHECK9-NEXT:    [[SUB30:%.*]] = sub nsw i32 [[DIV29]], 1
+// CHECK9-NEXT:    store i32 [[SUB30]], i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK9-NEXT:    [[TMP102:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_27]], align 4
+// CHECK9-NEXT:    [[ADD31:%.*]] = add nsw i32 [[TMP102]], 1
+// CHECK9-NEXT:    [[TMP103:%.*]] = zext i32 [[ADD31]] to i64
+// CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP103]])
+// CHECK9-NEXT:    [[KERNEL_ARGS32:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP104:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP104]], align 4
+// CHECK9-NEXT:    [[TMP105:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 3, i32* [[TMP105]], align 4
+// CHECK9-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP97]], i8*** [[TMP106]], align 8
+// CHECK9-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP98]], i8*** [[TMP107]], align 8
+// CHECK9-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* [[TMP99]], i64** [[TMP108]], align 8
+// CHECK9-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP109]], align 8
+// CHECK9-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP110]], align 8
+// CHECK9-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP111]], align 8
+// CHECK9-NEXT:    [[TMP112:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS32]])
+// CHECK9-NEXT:    [[TMP113:%.*]] = icmp ne i32 [[TMP112]], 0
+// CHECK9-NEXT:    br i1 [[TMP113]], label [[OMP_OFFLOAD_FAILED33:%.*]], label [[OMP_OFFLOAD_CONT34:%.*]]
+// CHECK9:       omp_offload.failed33:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110(i64 [[TMP78]], i64 [[TMP1]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT34]]
+// CHECK9:       omp_offload.cont34:
+// CHECK9-NEXT:    [[TMP114:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiLi10EEiT_(i32 noundef signext [[TMP114]])
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    [[TMP91:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP91]])
-// CHECK9-NEXT:    [[TMP92:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP92]]
+// CHECK9-NEXT:    [[TMP115:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK9-NEXT:    call void @llvm.stackrestore(i8* [[TMP115]])
+// CHECK9-NEXT:    [[TMP116:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP116]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100
@@ -1736,10 +1889,10 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[DOTOFFLOAD_PTRS2:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 8
 // CHECK9-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
-// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [1 x i8*], align 8
-// CHECK9-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK9-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [1 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [1 x i8*], align 8
+// CHECK9-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [1 x i8*], align 8
+// CHECK9-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK9-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK9-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP1:%.*]] = bitcast i8** [[TMP0]] to [10 x i32]**
@@ -1752,49 +1905,100 @@ int main (int argc, char **argv) {
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l79.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l79.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l79([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 8
-// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 8
-// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP13]], align 8
-// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 8
+// CHECK9-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 8
+// CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP21]], align 8
+// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK9-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK9:       omp_offload.failed5:
+// CHECK9-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK9-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK9-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK9:       omp_offload.failed6:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK9:       omp_offload.cont6:
-// CHECK9-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP19]], align 8
-// CHECK9-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [10 x i32]**
-// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP21]], align 8
-// CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP22]], align 8
-// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK9:       omp_offload.cont7:
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
+// CHECK9-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK9-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK9-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK9:       omp_offload.failed11:
+// CHECK9-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 8
+// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP46]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK9-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK9-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK9:       omp_offload.failed13:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK9:       omp_offload.cont12:
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK9:       omp_offload.cont14:
 // CHECK9-NEXT:    ret i32 0
 //
 //
@@ -2099,14 +2303,14 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[_TMP8:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_10:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[N_CASTED17:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS18:%.*]] = alloca [3 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS19:%.*]] = alloca [3 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS20:%.*]] = alloca [3 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES21:%.*]] = alloca [3 x i64], align 4
-// CHECK11-NEXT:    [[_TMP22:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_23:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[N_CASTED18:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS19:%.*]] = alloca [3 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS20:%.*]] = alloca [3 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS21:%.*]] = alloca [3 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_SIZES22:%.*]] = alloca [3 x i64], align 4
+// CHECK11-NEXT:    [[_TMP23:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    [[DOTCAPTURE_EXPR_24:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTCAPTURE_EXPR_25:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    store i8** [[ARGV]], i8*** [[ARGV_ADDR]], align 4
@@ -2163,128 +2367,179 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP28]], 1
 // CHECK11-NEXT:    [[TMP29:%.*]] = zext i32 [[ADD]] to i64
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 [[TMP29]])
-// CHECK11-NEXT:    [[TMP30:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, i32 3, i8** [[TMP23]], i8** [[TMP24]], i64* [[TMP25]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP31:%.*]] = icmp ne i32 [[TMP30]], 0
-// CHECK11-NEXT:    br i1 [[TMP31]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP24]], i8*** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP25]], i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP39:%.*]] = icmp ne i32 [[TMP38]], 0
+// CHECK11-NEXT:    br i1 [[TMP39]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100(i32 [[TMP3]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP32]], i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP33:%.*]] = load i32, i32* [[N_CASTED3]], align 4
-// CHECK11-NEXT:    [[TMP34:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = sext i32 [[TMP34]] to i64
-// CHECK11-NEXT:    [[TMP36:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP36]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i32 24, i1 false)
-// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = bitcast i8** [[TMP37]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP38]], align 4
-// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP40:%.*]] = bitcast i8** [[TMP39]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP33]], i32* [[TMP40]], align 4
-// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP41]], align 4
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP43:%.*]] = bitcast i8** [[TMP42]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP43]], align 4
-// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP45]], align 4
-// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP46]], align 4
-// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP48]], align 4
-// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP50:%.*]] = bitcast i8** [[TMP49]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP50]], align 4
-// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP35]], i64* [[TMP51]], align 4
-// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP52]], align 4
-// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP56:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP56]], i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[TMP57:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP57]], 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP40]], i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = load i32, i32* [[N_CASTED3]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = sext i32 [[TMP42]] to i64
+// CHECK11-NEXT:    [[TMP44:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES7]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP44]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.2 to i8*), i32 24, i1 false)
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP46:%.*]] = bitcast i8** [[TMP45]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP48:%.*]] = bitcast i8** [[TMP47]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP41]], i32* [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP49]], align 4
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP51:%.*]] = bitcast i8** [[TMP50]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP53:%.*]] = bitcast i8** [[TMP52]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP56:%.*]] = bitcast i8** [[TMP55]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP56]], align 4
+// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP58:%.*]] = bitcast i8** [[TMP57]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP58]], align 4
+// CHECK11-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP43]], i64* [[TMP59]], align 4
+// CHECK11-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP60]], align 4
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES7]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP64:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP64]], i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[TMP65:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK11-NEXT:    [[SUB11:%.*]] = sub nsw i32 [[TMP65]], 0
 // CHECK11-NEXT:    [[DIV12:%.*]] = sdiv i32 [[SUB11]], 1
 // CHECK11-NEXT:    [[SUB13:%.*]] = sub nsw i32 [[DIV12]], 1
 // CHECK11-NEXT:    store i32 [[SUB13]], i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[TMP58:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
-// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP58]], 1
-// CHECK11-NEXT:    [[TMP59:%.*]] = zext i32 [[ADD14]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP59]])
-// CHECK11-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105.region_id, i32 3, i8** [[TMP53]], i8** [[TMP54]], i64* [[TMP55]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK11-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED15:%.*]], label [[OMP_OFFLOAD_CONT16:%.*]]
-// CHECK11:       omp_offload.failed15:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105(i32 [[TMP33]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT16]]
-// CHECK11:       omp_offload.cont16:
-// CHECK11-NEXT:    [[TMP62:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP62]], i32* [[N_CASTED17]], align 4
-// CHECK11-NEXT:    [[TMP63:%.*]] = load i32, i32* [[N_CASTED17]], align 4
-// CHECK11-NEXT:    [[TMP64:%.*]] = mul nuw i32 [[TMP0]], 4
-// CHECK11-NEXT:    [[TMP65:%.*]] = sext i32 [[TMP64]] to i64
-// CHECK11-NEXT:    [[TMP66:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES21]] to i8*
-// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP66]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.5 to i8*), i32 24, i1 false)
-// CHECK11-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP68:%.*]] = bitcast i8** [[TMP67]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP63]], i32* [[TMP68]], align 4
-// CHECK11-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP70:%.*]] = bitcast i8** [[TMP69]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP63]], i32* [[TMP70]], align 4
-// CHECK11-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP71]], align 4
-// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP73:%.*]] = bitcast i8** [[TMP72]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP73]], align 4
-// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 1
-// CHECK11-NEXT:    [[TMP75:%.*]] = bitcast i8** [[TMP74]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP75]], align 4
-// CHECK11-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 1
-// CHECK11-NEXT:    store i8* null, i8** [[TMP76]], align 4
-// CHECK11-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP78:%.*]] = bitcast i8** [[TMP77]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP78]], align 4
-// CHECK11-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 2
-// CHECK11-NEXT:    [[TMP80:%.*]] = bitcast i8** [[TMP79]] to i32**
-// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP80]], align 4
-// CHECK11-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 2
-// CHECK11-NEXT:    store i64 [[TMP65]], i64* [[TMP81]], align 4
-// CHECK11-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS20]], i32 0, i32 2
-// CHECK11-NEXT:    store i8* null, i8** [[TMP82]], align 4
-// CHECK11-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS18]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS19]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES21]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP86:%.*]] = load i32, i32* [[N]], align 4
-// CHECK11-NEXT:    store i32 [[TMP86]], i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK11-NEXT:    [[TMP87:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_23]], align 4
-// CHECK11-NEXT:    [[SUB25:%.*]] = sub nsw i32 [[TMP87]], 0
-// CHECK11-NEXT:    [[DIV26:%.*]] = sdiv i32 [[SUB25]], 1
-// CHECK11-NEXT:    [[SUB27:%.*]] = sub nsw i32 [[DIV26]], 1
-// CHECK11-NEXT:    store i32 [[SUB27]], i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[TMP88:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
-// CHECK11-NEXT:    [[ADD28:%.*]] = add nsw i32 [[TMP88]], 1
-// CHECK11-NEXT:    [[TMP89:%.*]] = zext i32 [[ADD28]] to i64
-// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP89]])
-// CHECK11-NEXT:    [[TMP90:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110.region_id, i32 3, i8** [[TMP83]], i8** [[TMP84]], i64* [[TMP85]], i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP91:%.*]] = icmp ne i32 [[TMP90]], 0
-// CHECK11-NEXT:    br i1 [[TMP91]], label [[OMP_OFFLOAD_FAILED29:%.*]], label [[OMP_OFFLOAD_CONT30:%.*]]
-// CHECK11:       omp_offload.failed29:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110(i32 [[TMP63]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT30]]
-// CHECK11:       omp_offload.cont30:
-// CHECK11-NEXT:    [[TMP92:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
-// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP92]])
+// CHECK11-NEXT:    [[TMP66:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_10]], align 4
+// CHECK11-NEXT:    [[ADD14:%.*]] = add nsw i32 [[TMP66]], 1
+// CHECK11-NEXT:    [[TMP67:%.*]] = zext i32 [[ADD14]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP67]])
+// CHECK11-NEXT:    [[KERNEL_ARGS15:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP68:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP68]], align 4
+// CHECK11-NEXT:    [[TMP69:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP69]], align 4
+// CHECK11-NEXT:    [[TMP70:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP61]], i8*** [[TMP70]], align 4
+// CHECK11-NEXT:    [[TMP71:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP62]], i8*** [[TMP71]], align 4
+// CHECK11-NEXT:    [[TMP72:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP63]], i64** [[TMP72]], align 4
+// CHECK11-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP73]], align 4
+// CHECK11-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP74]], align 4
+// CHECK11-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP75]], align 4
+// CHECK11-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS15]])
+// CHECK11-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
+// CHECK11-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED16:%.*]], label [[OMP_OFFLOAD_CONT17:%.*]]
+// CHECK11:       omp_offload.failed16:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l105(i32 [[TMP41]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT17]]
+// CHECK11:       omp_offload.cont17:
+// CHECK11-NEXT:    [[TMP78:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP78]], i32* [[N_CASTED18]], align 4
+// CHECK11-NEXT:    [[TMP79:%.*]] = load i32, i32* [[N_CASTED18]], align 4
+// CHECK11-NEXT:    [[TMP80:%.*]] = mul nuw i32 [[TMP0]], 4
+// CHECK11-NEXT:    [[TMP81:%.*]] = sext i32 [[TMP80]] to i64
+// CHECK11-NEXT:    [[TMP82:%.*]] = bitcast [3 x i64]* [[DOTOFFLOAD_SIZES22]] to i8*
+// CHECK11-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 [[TMP82]], i8* align 4 bitcast ([3 x i64]* @.offload_sizes.5 to i8*), i32 24, i1 false)
+// CHECK11-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP84:%.*]] = bitcast i8** [[TMP83]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP79]], i32* [[TMP84]], align 4
+// CHECK11-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP86:%.*]] = bitcast i8** [[TMP85]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP79]], i32* [[TMP86]], align 4
+// CHECK11-NEXT:    [[TMP87:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP87]], align 4
+// CHECK11-NEXT:    [[TMP88:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP89:%.*]] = bitcast i8** [[TMP88]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP89]], align 4
+// CHECK11-NEXT:    [[TMP90:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 1
+// CHECK11-NEXT:    [[TMP91:%.*]] = bitcast i8** [[TMP90]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP0]], i32* [[TMP91]], align 4
+// CHECK11-NEXT:    [[TMP92:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 1
+// CHECK11-NEXT:    store i8* null, i8** [[TMP92]], align 4
+// CHECK11-NEXT:    [[TMP93:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP94:%.*]] = bitcast i8** [[TMP93]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP94]], align 4
+// CHECK11-NEXT:    [[TMP95:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 2
+// CHECK11-NEXT:    [[TMP96:%.*]] = bitcast i8** [[TMP95]] to i32**
+// CHECK11-NEXT:    store i32* [[VLA]], i32** [[TMP96]], align 4
+// CHECK11-NEXT:    [[TMP97:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 2
+// CHECK11-NEXT:    store i64 [[TMP81]], i64* [[TMP97]], align 4
+// CHECK11-NEXT:    [[TMP98:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_MAPPERS21]], i32 0, i32 2
+// CHECK11-NEXT:    store i8* null, i8** [[TMP98]], align 4
+// CHECK11-NEXT:    [[TMP99:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_BASEPTRS19]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP100:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[DOTOFFLOAD_PTRS20]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP101:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[DOTOFFLOAD_SIZES22]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP102:%.*]] = load i32, i32* [[N]], align 4
+// CHECK11-NEXT:    store i32 [[TMP102]], i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK11-NEXT:    [[TMP103:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_24]], align 4
+// CHECK11-NEXT:    [[SUB26:%.*]] = sub nsw i32 [[TMP103]], 0
+// CHECK11-NEXT:    [[DIV27:%.*]] = sdiv i32 [[SUB26]], 1
+// CHECK11-NEXT:    [[SUB28:%.*]] = sub nsw i32 [[DIV27]], 1
+// CHECK11-NEXT:    store i32 [[SUB28]], i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[TMP104:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_25]], align 4
+// CHECK11-NEXT:    [[ADD29:%.*]] = add nsw i32 [[TMP104]], 1
+// CHECK11-NEXT:    [[TMP105:%.*]] = zext i32 [[ADD29]] to i64
+// CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 [[TMP105]])
+// CHECK11-NEXT:    [[KERNEL_ARGS30:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP106:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP106]], align 4
+// CHECK11-NEXT:    [[TMP107:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 3, i32* [[TMP107]], align 4
+// CHECK11-NEXT:    [[TMP108:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP99]], i8*** [[TMP108]], align 4
+// CHECK11-NEXT:    [[TMP109:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP100]], i8*** [[TMP109]], align 4
+// CHECK11-NEXT:    [[TMP110:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* [[TMP101]], i64** [[TMP110]], align 4
+// CHECK11-NEXT:    [[TMP111:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([3 x i64], [3 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP111]], align 4
+// CHECK11-NEXT:    [[TMP112:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP112]], align 4
+// CHECK11-NEXT:    [[TMP113:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP113]], align 4
+// CHECK11-NEXT:    [[TMP114:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS30]])
+// CHECK11-NEXT:    [[TMP115:%.*]] = icmp ne i32 [[TMP114]], 0
+// CHECK11-NEXT:    br i1 [[TMP115]], label [[OMP_OFFLOAD_FAILED31:%.*]], label [[OMP_OFFLOAD_CONT32:%.*]]
+// CHECK11:       omp_offload.failed31:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l110(i32 [[TMP79]], i32 [[TMP0]], i32* [[VLA]]) #[[ATTR3]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT32]]
+// CHECK11:       omp_offload.cont32:
+// CHECK11-NEXT:    [[TMP116:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4
+// CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiLi10EEiT_(i32 noundef [[TMP116]])
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    [[TMP93:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP93]])
-// CHECK11-NEXT:    [[TMP94:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP94]]
+// CHECK11-NEXT:    [[TMP117:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK11-NEXT:    call void @llvm.stackrestore(i8* [[TMP117]])
+// CHECK11-NEXT:    [[TMP118:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP118]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l100
@@ -2690,10 +2945,10 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[DOTOFFLOAD_PTRS2:%.*]] = alloca [1 x i8*], align 4
 // CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS3:%.*]] = alloca [1 x i8*], align 4
 // CHECK11-NEXT:    [[_TMP4:%.*]] = alloca i32, align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS7:%.*]] = alloca [1 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS8:%.*]] = alloca [1 x i8*], align 4
-// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS9:%.*]] = alloca [1 x i8*], align 4
-// CHECK11-NEXT:    [[_TMP10:%.*]] = alloca i32, align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_BASEPTRS8:%.*]] = alloca [1 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_PTRS9:%.*]] = alloca [1 x i8*], align 4
+// CHECK11-NEXT:    [[DOTOFFLOAD_MAPPERS10:%.*]] = alloca [1 x i8*], align 4
+// CHECK11-NEXT:    [[_TMP11:%.*]] = alloca i32, align 4
 // CHECK11-NEXT:    store i32 [[ARGC]], i32* [[ARGC_ADDR]], align 4
 // CHECK11-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP1:%.*]] = bitcast i8** [[TMP0]] to [10 x i32]**
@@ -2706,49 +2961,100 @@ int main (int argc, char **argv) {
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l79.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l79.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l79([10 x i32]* [[A]]) #[[ATTR3]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP10:%.*]] = bitcast i8** [[TMP9]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP10]], align 4
-// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP12:%.*]] = bitcast i8** [[TMP11]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP12]], align 4
-// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP13]], align 4
-// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP18:%.*]] = bitcast i8** [[TMP17]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP18]], align 4
+// CHECK11-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP20:%.*]] = bitcast i8** [[TMP19]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP20]], align 4
+// CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP21]], align 4
+// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, i32 1, i8** [[TMP14]], i8** [[TMP15]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
-// CHECK11-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK11:       omp_offload.failed5:
+// CHECK11-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP24]], align 4
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP25]], align 4
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP23]], i8*** [[TMP27]], align 4
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.11, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.12, i32 0, i32 0), i64** [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK11-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
+// CHECK11-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK11:       omp_offload.failed6:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l84([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK11:       omp_offload.cont6:
-// CHECK11-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP19:%.*]] = bitcast i8** [[TMP18]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP19]], align 4
-// CHECK11-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP21:%.*]] = bitcast i8** [[TMP20]] to [10 x i32]**
-// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP21]], align 4
-// CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS9]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP22]], align 4
-// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS7]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK11:       omp_offload.cont7:
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP35:%.*]] = bitcast i8** [[TMP34]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to [10 x i32]**
+// CHECK11-NEXT:    store [10 x i32]* [[A]], [10 x i32]** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS10]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS8]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS9]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 10)
-// CHECK11-NEXT:    [[TMP25:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, i32 1, i8** [[TMP23]], i8** [[TMP24]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP26:%.*]] = icmp ne i32 [[TMP25]], 0
-// CHECK11-NEXT:    br i1 [[TMP26]], label [[OMP_OFFLOAD_FAILED11:%.*]], label [[OMP_OFFLOAD_CONT12:%.*]]
-// CHECK11:       omp_offload.failed11:
+// CHECK11-NEXT:    [[KERNEL_ARGS12:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP40]], i8*** [[TMP44]], align 4
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.14, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.15, i32 0, i32 0), i64** [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS12]])
+// CHECK11-NEXT:    [[TMP50:%.*]] = icmp ne i32 [[TMP49]], 0
+// CHECK11-NEXT:    br i1 [[TMP50]], label [[OMP_OFFLOAD_FAILED13:%.*]], label [[OMP_OFFLOAD_CONT14:%.*]]
+// CHECK11:       omp_offload.failed13:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiLi10EEiT__l89([10 x i32]* [[A]]) #[[ATTR3]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT12]]
-// CHECK11:       omp_offload.cont12:
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT14]]
+// CHECK11:       omp_offload.cont14:
 // CHECK11-NEXT:    ret i32 0
 //
 //

diff --git a/clang/test/OpenMP/teams_distribute_simd_firstprivate_codegen.cpp b/clang/test/OpenMP/teams_distribute_simd_firstprivate_codegen.cpp
index 8213061..4a22333 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_simd_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_simd_firstprivate_codegen.cpp
@@ -317,9 +317,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK1-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP37]], align 8
+// CHECK1-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK1-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK1-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94(i64 [[TMP1]], [2 x i32]* @vec, [2 x %struct.S]* @s_arr, %struct.S* @var, i64 [[TMP3]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -600,27 +617,44 @@ int main() {
 // CHECK1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK1-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK1-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK1-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK1-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK1-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK1-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP32]]
+// CHECK1-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN2StC2Ev
@@ -1121,9 +1155,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 5, i8** [[TMP29]], i8** [[TMP30]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
-// CHECK3-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 5, i32* [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP29]], i8*** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP37]], align 4
+// CHECK3-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK3-NEXT:    [[TMP39:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP40:%.*]] = icmp ne i32 [[TMP39]], 0
+// CHECK3-NEXT:    br i1 [[TMP40]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94(i32 [[TMP1]], [2 x i32]* @vec, [2 x %struct.S]* @s_arr, %struct.S* @var, i32 [[TMP3]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1395,27 +1446,44 @@ int main() {
 // CHECK3-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK3-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK3-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK3-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK3-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK3-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK3-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK3-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK3-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK3-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP32]]
+// CHECK3-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN2StC2Ev

diff --git a/clang/test/OpenMP/teams_distribute_simd_lastprivate_codegen.cpp b/clang/test/OpenMP/teams_distribute_simd_lastprivate_codegen.cpp
index 8cb0f49..7e98688 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_simd_lastprivate_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_simd_lastprivate_codegen.cpp
@@ -599,9 +599,26 @@ int main() {
 // CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK9-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 8
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 8
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK9-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i64 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -609,18 +626,18 @@ int main() {
 // CHECK9-NEXT:    [[CALL:%.*]] = call noundef signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE3:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done3:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP39]]
+// CHECK9-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -912,27 +929,44 @@ int main() {
 // CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK9-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK9-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK9-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK9-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 8
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP33]], align 8
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK9-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i64 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done2:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP32]]
+// CHECK9-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1306,9 +1340,26 @@ int main() {
 // CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP36:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96.region_id, i32 5, i8** [[TMP34]], i8** [[TMP35]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP37:%.*]] = icmp ne i32 [[TMP36]], 0
-// CHECK11-NEXT:    br i1 [[TMP37]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP34]], i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP35]], i8*** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP45:%.*]] = icmp ne i32 [[TMP44]], 0
+// CHECK11-NEXT:    br i1 [[TMP45]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l96(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[TMP4]], i32 [[TMP6]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1316,18 +1367,18 @@ int main() {
 // CHECK11-NEXT:    [[CALL:%.*]] = call noundef i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP38]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP46]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP39:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP39]]
+// CHECK11-NEXT:    [[TMP47:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP47]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1614,27 +1665,44 @@ int main() {
 // CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK11-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK11-NEXT:    [[TMP29:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, i32 4, i8** [[TMP27]], i8** [[TMP28]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK11-NEXT:    [[TMP30:%.*]] = icmp ne i32 [[TMP29]], 0
-// CHECK11-NEXT:    br i1 [[TMP30]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP27]], i8*** [[TMP31]], align 4
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP28]], i8*** [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP38:%.*]] = icmp ne i32 [[TMP37]], 0
+// CHECK11-NEXT:    br i1 [[TMP38]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l49(i32 [[TMP3]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[TMP4]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP31]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP39]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done2:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP32:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP32]]
+// CHECK11-NEXT:    [[TMP40:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP40]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev

diff --git a/clang/test/OpenMP/teams_distribute_simd_private_codegen.cpp b/clang/test/OpenMP/teams_distribute_simd_private_codegen.cpp
index 28ce249..583fd19 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_simd_private_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_simd_private_codegen.cpp
@@ -243,9 +243,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK1-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP2]], align 8
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -396,27 +413,44 @@ int main() {
 // CHECK1-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 8
 // CHECK1-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 8
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK1-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK1-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56() #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK1:       omp_offload.cont:
 // CHECK1-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK1-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK1-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK1:       arraydestroy.body:
-// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK1-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK1-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK1-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK1:       arraydestroy.done2:
 // CHECK1-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK1-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK1-NEXT:    ret i32 [[TMP4]]
+// CHECK1-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK1-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK1-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev
@@ -745,9 +779,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP0:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
-// CHECK3-NEXT:    br i1 [[TMP1]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP0:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP0]], align 4
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l94() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -896,27 +947,44 @@ int main() {
 // CHECK3-NEXT:    store %struct.S.0* [[TEST]], %struct.S.0** [[VAR]], align 4
 // CHECK3-NEXT:    store %struct.S.0* undef, %struct.S.0** [[_TMP1]], align 4
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB2]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK3-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK3-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK3-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK3-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB2]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l56() #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK3:       omp_offload.cont:
 // CHECK3-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK3-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK3-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK3:       arraydestroy.body:
-// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK3-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR2]]
 // CHECK3-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK3-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK3:       arraydestroy.done2:
 // CHECK3-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR2]]
-// CHECK3-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK3-NEXT:    ret i32 [[TMP4]]
+// CHECK3-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK3-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK3-LABEL: define {{[^@]+}}@_ZN1SIiEC1Ev

diff --git a/clang/test/OpenMP/teams_distribute_simd_reduction_codegen.cpp b/clang/test/OpenMP/teams_distribute_simd_reduction_codegen.cpp
index 8fdf7c6..d9079b4 100644 (file)
--- a/clang/test/OpenMP/teams_distribute_simd_reduction_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_simd_reduction_codegen.cpp
@@ -107,9 +107,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l63.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK1-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l63.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK1-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l63(i64 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -279,9 +296,26 @@ int main() {
 // CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK1-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK1-NEXT:    [[TMP10:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK1-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[TMP10]], 0
-// CHECK1-NEXT:    br i1 [[TMP11]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP9]], i8*** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP16]], align 8
+// CHECK1-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP17]], align 8
+// CHECK1-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK1-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i64 [[TMP2]]) #[[ATTR2]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -453,9 +487,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3:[0-9]+]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l63.region_id, i32 1, i8** [[TMP7]], i8** [[TMP8]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
-// CHECK3-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l63.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP18:%.*]] = icmp ne i32 [[TMP17]], 0
+// CHECK3-NEXT:    br i1 [[TMP18]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l63(i32 [[TMP1]]) #[[ATTR2:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -623,9 +674,26 @@ int main() {
 // CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK3-NEXT:    call void @__kmpc_push_target_tripcount_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i64 2)
-// CHECK3-NEXT:    [[TMP10:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB3]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, i32 1, i8** [[TMP8]], i8** [[TMP9]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i8** null, i8** null, i32 0, i32 1)
-// CHECK3-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[TMP10]], 0
-// CHECK3-NEXT:    br i1 [[TMP11]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP8]], i8*** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP9]], i8*** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.3, i32 0, i32 0), i64** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.4, i32 0, i32 0), i64** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP16]], align 4
+// CHECK3-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP17]], align 4
+// CHECK3-NEXT:    [[TMP18:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB3]], i64 -1, i32 0, i32 1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP19:%.*]] = icmp ne i32 [[TMP18]], 0
+// CHECK3-NEXT:    br i1 [[TMP19]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l32(i32 [[TMP2]]) #[[ATTR2]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/clang/test/OpenMP/teams_firstprivate_codegen.cpp b/clang/test/OpenMP/teams_firstprivate_codegen.cpp
index 323ecbe..328b224 100644 (file)
--- a/clang/test/OpenMP/teams_firstprivate_codegen.cpp
+++ b/clang/test/OpenMP/teams_firstprivate_codegen.cpp
@@ -362,50 +362,84 @@ void array_func(float a[3], St s[2], int n, long double vla1[n]) {
 // CHECK9-NEXT:    store i8* null, i8** [[TMP29]], align 8
 // CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l109.region_id, i32 5, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK9-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 5, i32* [[TMP33]], align 4
+// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP31]], i8*** [[TMP35]], align 8
+// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP38]], align 8
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP39]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l109.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK9-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l109(i64 [[TMP2]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[VAR]], i64 [[TMP4]]) #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP34:%.*]] = load i32, i32* [[T_VAR]], align 4
+// CHECK9-NEXT:    [[TMP42:%.*]] = load i32, i32* [[T_VAR]], align 4
 // CHECK9-NEXT:    [[CONV3:%.*]] = bitcast i64* [[T_VAR_CASTED2]] to i32*
-// CHECK9-NEXT:    store i32 [[TMP34]], i32* [[CONV3]], align 4
-// CHECK9-NEXT:    [[TMP35:%.*]] = load i64, i64* [[T_VAR_CASTED2]], align 8
-// CHECK9-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP35]], i64* [[TMP37]], align 8
-// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i64*
-// CHECK9-NEXT:    store i64 [[TMP35]], i64* [[TMP39]], align 8
-// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP40]], align 8
-// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK9-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED7:%.*]], label [[OMP_OFFLOAD_CONT8:%.*]]
-// CHECK9:       omp_offload.failed7:
-// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116(i64 [[TMP35]]) #[[ATTR4]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT8]]
-// CHECK9:       omp_offload.cont8:
+// CHECK9-NEXT:    store i32 [[TMP42]], i32* [[CONV3]], align 4
+// CHECK9-NEXT:    [[TMP43:%.*]] = load i64, i64* [[T_VAR_CASTED2]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP43]], i64* [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i64*
+// CHECK9-NEXT:    store i64 [[TMP43]], i64* [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS6]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP48]], align 8
+// CHECK9-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS4]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS5]], i32 0, i32 0
+// CHECK9-NEXT:    [[KERNEL_ARGS7:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP51]], align 4
+// CHECK9-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP52]], align 4
+// CHECK9-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP49]], i8*** [[TMP53]], align 8
+// CHECK9-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP50]], i8*** [[TMP54]], align 8
+// CHECK9-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP55]], align 8
+// CHECK9-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP56]], align 8
+// CHECK9-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP57]], align 8
+// CHECK9-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP58]], align 8
+// CHECK9-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS7]])
+// CHECK9-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
+// CHECK9-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED8:%.*]], label [[OMP_OFFLOAD_CONT9:%.*]]
+// CHECK9:       omp_offload.failed8:
+// CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116(i64 [[TMP43]]) #[[ATTR4]]
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT9]]
+// CHECK9:       omp_offload.cont9:
 // CHECK9-NEXT:    [[CALL:%.*]] = call signext i32 @_Z5tmainIiET_v()
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* nonnull align 4 dereferenceable(4) [[VAR]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP45]], [[OMP_OFFLOAD_CONT8]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP61]], [[OMP_OFFLOAD_CONT9]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
-// CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE9:%.*]], label [[ARRAYDESTROY_BODY]]
-// CHECK9:       arraydestroy.done9:
+// CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE10:%.*]], label [[ARRAYDESTROY_BODY]]
+// CHECK9:       arraydestroy.done10:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP46:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP46]]
+// CHECK9-NEXT:    [[TMP62:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP62]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -663,45 +697,79 @@ void array_func(float a[3], St s[2], int n, long double vla1[n]) {
 // CHECK9-NEXT:    store i8* null, i8** [[TMP20]], align 8
 // CHECK9-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP23:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75.region_id, i32 4, i8** [[TMP21]], i8** [[TMP22]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP24:%.*]] = icmp ne i32 [[TMP23]], 0
-// CHECK9-NEXT:    br i1 [[TMP24]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK9-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 4, i32* [[TMP24]], align 4
+// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 8
+// CHECK9-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 8
+// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP27]], align 8
+// CHECK9-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP28]], align 8
+// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP29]], align 8
+// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP30]], align 8
+// CHECK9-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
+// CHECK9-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75(i32* [[T_VAR]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[VAR]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK9:       omp_offload.cont:
-// CHECK9-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i32**
-// CHECK9-NEXT:    store i32* [[T_VAR]], i32** [[TMP26]], align 8
-// CHECK9-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i32**
-// CHECK9-NEXT:    store i32* [[T_VAR]], i32** [[TMP28]], align 8
-// CHECK9-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
-// CHECK9-NEXT:    store i8* null, i8** [[TMP29]], align 8
-// CHECK9-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81.region_id, i32 1, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK9-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED4:%.*]], label [[OMP_OFFLOAD_CONT5:%.*]]
-// CHECK9:       omp_offload.failed4:
+// CHECK9-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP34:%.*]] = bitcast i8** [[TMP33]] to i32**
+// CHECK9-NEXT:    store i32* [[T_VAR]], i32** [[TMP34]], align 8
+// CHECK9-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to i32**
+// CHECK9-NEXT:    store i32* [[T_VAR]], i32** [[TMP36]], align 8
+// CHECK9-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i64 0, i64 0
+// CHECK9-NEXT:    store i8* null, i8** [[TMP37]], align 8
+// CHECK9-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK9-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK9-NEXT:    [[KERNEL_ARGS4:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK9-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK9-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK9-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP38]], i8*** [[TMP42]], align 8
+// CHECK9-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 8
+// CHECK9-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP44]], align 8
+// CHECK9-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP45]], align 8
+// CHECK9-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP46]], align 8
+// CHECK9-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP47]], align 8
+// CHECK9-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]])
+// CHECK9-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK9-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
+// CHECK9:       omp_offload.failed5:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81(i32* [[T_VAR]]) #[[ATTR4]]
-// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT5]]
-// CHECK9:       omp_offload.cont5:
+// CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
+// CHECK9:       omp_offload.cont6:
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* nonnull align 4 dereferenceable(4) [[VAR]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP34]], [[OMP_OFFLOAD_CONT5]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP50]], [[OMP_OFFLOAD_CONT6]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
-// CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE6:%.*]], label [[ARRAYDESTROY_BODY]]
-// CHECK9:       arraydestroy.done6:
+// CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE7:%.*]], label [[ARRAYDESTROY_BODY]]
+// CHECK9:       arraydestroy.done7:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP35:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP35]]
+// CHECK9-NEXT:    [[TMP51:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1091,49 +1159,83 @@ void array_func(float a[3], St s[2], int n, long double vla1[n]) {
 // CHECK11-NEXT:    store i8* null, i8** [[TMP29]], align 4
 // CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l109.region_id, i32 5, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK11-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP32:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP32]], align 4
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 5, i32* [[TMP33]], align 4
+// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP30]], i8*** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP31]], i8*** [[TMP35]], align 4
+// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([5 x i64], [5 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP38]], align 4
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP39]], align 4
+// CHECK11-NEXT:    [[TMP40:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l109.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP41:%.*]] = icmp ne i32 [[TMP40]], 0
+// CHECK11-NEXT:    br i1 [[TMP41]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l109(i32 [[TMP2]], [2 x i32]* [[VEC]], [2 x %struct.S]* [[S_ARR]], %struct.S* [[VAR]], i32 [[TMP4]]) #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP34:%.*]] = load i32, i32* [[T_VAR]], align 4
-// CHECK11-NEXT:    store i32 [[TMP34]], i32* [[T_VAR_CASTED1]], align 4
-// CHECK11-NEXT:    [[TMP35:%.*]] = load i32, i32* [[T_VAR_CASTED1]], align 4
-// CHECK11-NEXT:    [[TMP36:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP37:%.*]] = bitcast i8** [[TMP36]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP35]], i32* [[TMP37]], align 4
-// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP39:%.*]] = bitcast i8** [[TMP38]] to i32*
-// CHECK11-NEXT:    store i32 [[TMP35]], i32* [[TMP39]], align 4
-// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP40]], align 4
-// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP43:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116.region_id, i32 1, i8** [[TMP41]], i8** [[TMP42]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP44:%.*]] = icmp ne i32 [[TMP43]], 0
-// CHECK11-NEXT:    br i1 [[TMP44]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
-// CHECK11:       omp_offload.failed5:
-// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116(i32 [[TMP35]]) #[[ATTR4]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
-// CHECK11:       omp_offload.cont6:
+// CHECK11-NEXT:    [[TMP42:%.*]] = load i32, i32* [[T_VAR]], align 4
+// CHECK11-NEXT:    store i32 [[TMP42]], i32* [[T_VAR_CASTED1]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = load i32, i32* [[T_VAR_CASTED1]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP45:%.*]] = bitcast i8** [[TMP44]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP43]], i32* [[TMP45]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP47:%.*]] = bitcast i8** [[TMP46]] to i32*
+// CHECK11-NEXT:    store i32 [[TMP43]], i32* [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS4]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP48]], align 4
+// CHECK11-NEXT:    [[TMP49:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS3]], i32 0, i32 0
+// CHECK11-NEXT:    [[KERNEL_ARGS5:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP51:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP51]], align 4
+// CHECK11-NEXT:    [[TMP52:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP52]], align 4
+// CHECK11-NEXT:    [[TMP53:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP49]], i8*** [[TMP53]], align 4
+// CHECK11-NEXT:    [[TMP54:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP50]], i8*** [[TMP54]], align 4
+// CHECK11-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.2, i32 0, i32 0), i64** [[TMP55]], align 4
+// CHECK11-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP56]], align 4
+// CHECK11-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP57]], align 4
+// CHECK11-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP58]], align 4
+// CHECK11-NEXT:    [[TMP59:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS5]])
+// CHECK11-NEXT:    [[TMP60:%.*]] = icmp ne i32 [[TMP59]], 0
+// CHECK11-NEXT:    br i1 [[TMP60]], label [[OMP_OFFLOAD_FAILED6:%.*]], label [[OMP_OFFLOAD_CONT7:%.*]]
+// CHECK11:       omp_offload.failed6:
+// CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l116(i32 [[TMP43]]) #[[ATTR4]]
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT7]]
+// CHECK11:       omp_offload.cont7:
 // CHECK11-NEXT:    [[CALL:%.*]] = call i32 @_Z5tmainIiET_v()
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* nonnull align 4 dereferenceable(4) [[VAR]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP45]], [[OMP_OFFLOAD_CONT6]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP61]], [[OMP_OFFLOAD_CONT7]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
-// CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE7:%.*]], label [[ARRAYDESTROY_BODY]]
-// CHECK11:       arraydestroy.done7:
+// CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE8:%.*]], label [[ARRAYDESTROY_BODY]]
+// CHECK11:       arraydestroy.done8:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP46:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP46]]
+// CHECK11-NEXT:    [[TMP62:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP62]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC1Ev
@@ -1382,45 +1484,79 @@ void array_func(float a[3], St s[2], int n, long double vla1[n]) {
 // CHECK11-NEXT:    store i8* null, i8** [[TMP20]], align 4
 // CHECK11-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP22:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP23:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75.region_id, i32 4, i8** [[TMP21]], i8** [[TMP22]], i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP24:%.*]] = icmp ne i32 [[TMP23]], 0
-// CHECK11-NEXT:    br i1 [[TMP24]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP23:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP23]], align 4
+// CHECK11-NEXT:    [[TMP24:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 4, i32* [[TMP24]], align 4
+// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP21]], i8*** [[TMP25]], align 4
+// CHECK11-NEXT:    [[TMP26:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP22]], i8*** [[TMP26]], align 4
+// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_sizes.5, i32 0, i32 0), i64** [[TMP27]], align 4
+// CHECK11-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([4 x i64], [4 x i64]* @.offload_maptypes.6, i32 0, i32 0), i64** [[TMP28]], align 4
+// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP29]], align 4
+// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP30]], align 4
+// CHECK11-NEXT:    [[TMP31:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP32:%.*]] = icmp ne i32 [[TMP31]], 0
+// CHECK11-NEXT:    br i1 [[TMP32]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l75(i32* [[T_VAR]], [2 x i32]* [[VEC]], [2 x %struct.S.0]* [[S_ARR]], %struct.S.0* [[VAR]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK11:       omp_offload.cont:
-// CHECK11-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP26:%.*]] = bitcast i8** [[TMP25]] to i32**
-// CHECK11-NEXT:    store i32* [[T_VAR]], i32** [[TMP26]], align 4
-// CHECK11-NEXT:    [[TMP27:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP28:%.*]] = bitcast i8** [[TMP27]] to i32**
-// CHECK11-NEXT:    store i32* [[T_VAR]], i32** [[TMP28]], align 4
-// CHECK11-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
-// CHECK11-NEXT:    store i8* null, i8** [[TMP29]], align 4
-// CHECK11-NEXT:    [[TMP30:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP31:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP32:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81.region_id, i32 1, i8** [[TMP30]], i8** [[TMP31]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP33:%.*]] = icmp ne i32 [[TMP32]], 0
-// CHECK11-NEXT:    br i1 [[TMP33]], label [[OMP_OFFLOAD_FAILED4:%.*]], label [[OMP_OFFLOAD_CONT5:%.*]]
-// CHECK11:       omp_offload.failed4:
+// CHECK11-NEXT:    [[TMP33:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP34:%.*]] = bitcast i8** [[TMP33]] to i32**
+// CHECK11-NEXT:    store i32* [[T_VAR]], i32** [[TMP34]], align 4
+// CHECK11-NEXT:    [[TMP35:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP36:%.*]] = bitcast i8** [[TMP35]] to i32**
+// CHECK11-NEXT:    store i32* [[T_VAR]], i32** [[TMP36]], align 4
+// CHECK11-NEXT:    [[TMP37:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS3]], i32 0, i32 0
+// CHECK11-NEXT:    store i8* null, i8** [[TMP37]], align 4
+// CHECK11-NEXT:    [[TMP38:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS1]], i32 0, i32 0
+// CHECK11-NEXT:    [[TMP39:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS2]], i32 0, i32 0
+// CHECK11-NEXT:    [[KERNEL_ARGS4:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS]], align 8
+// CHECK11-NEXT:    [[TMP40:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP40]], align 4
+// CHECK11-NEXT:    [[TMP41:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP41]], align 4
+// CHECK11-NEXT:    [[TMP42:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP38]], i8*** [[TMP42]], align 4
+// CHECK11-NEXT:    [[TMP43:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP39]], i8*** [[TMP43]], align 4
+// CHECK11-NEXT:    [[TMP44:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.8, i32 0, i32 0), i64** [[TMP44]], align 4
+// CHECK11-NEXT:    [[TMP45:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.9, i32 0, i32 0), i64** [[TMP45]], align 4
+// CHECK11-NEXT:    [[TMP46:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP46]], align 4
+// CHECK11-NEXT:    [[TMP47:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP47]], align 4
+// CHECK11-NEXT:    [[TMP48:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS4]])
+// CHECK11-NEXT:    [[TMP49:%.*]] = icmp ne i32 [[TMP48]], 0
+// CHECK11-NEXT:    br i1 [[TMP49]], label [[OMP_OFFLOAD_FAILED5:%.*]], label [[OMP_OFFLOAD_CONT6:%.*]]
+// CHECK11:       omp_offload.failed5:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l81(i32* [[T_VAR]]) #[[ATTR4]]
-// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT5]]
-// CHECK11:       omp_offload.cont5:
+// CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT6]]
+// CHECK11:       omp_offload.cont6:
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* nonnull align 4 dereferenceable(4) [[VAR]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP50:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP34]], [[OMP_OFFLOAD_CONT5]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP50]], [[OMP_OFFLOAD_CONT6]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
-// CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE6:%.*]], label [[ARRAYDESTROY_BODY]]
-// CHECK11:       arraydestroy.done6:
+// CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE7:%.*]], label [[ARRAYDESTROY_BODY]]
+// CHECK11:       arraydestroy.done7:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP35:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP35]]
+// CHECK11-NEXT:    [[TMP51:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP51]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN1SIfEC2Ev
@@ -1845,15 +1981,32 @@ void array_func(float a[3], St s[2], int n, long double vla1[n]) {
 // CHECK17-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [8 x i8*], [8 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [8 x i8*], [8 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [8 x i64], [8 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP60:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z10array_funcPfP2StiPg_l152.region_id, i32 8, i8** [[TMP57]], i8** [[TMP58]], i64* [[TMP59]], i64* getelementptr inbounds ([8 x i64], [8 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP61:%.*]] = icmp ne i32 [[TMP60]], 0
-// CHECK17-NEXT:    br i1 [[TMP61]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP60]], align 4
+// CHECK17-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 8, i32* [[TMP61]], align 4
+// CHECK17-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP57]], i8*** [[TMP62]], align 8
+// CHECK17-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP58]], i8*** [[TMP63]], align 8
+// CHECK17-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP59]], i64** [[TMP64]], align 8
+// CHECK17-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([8 x i64], [8 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP65]], align 8
+// CHECK17-NEXT:    [[TMP66:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP66]], align 8
+// CHECK17-NEXT:    [[TMP67:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP67]], align 8
+// CHECK17-NEXT:    [[TMP68:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z10array_funcPfP2StiPg_l152.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP69:%.*]] = icmp ne i32 [[TMP68]], 0
+// CHECK17-NEXT:    br i1 [[TMP69]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z10array_funcPfP2StiPg_l152(float* [[TMP8]], %struct.St* [[TMP9]], i64 [[TMP1]], ppc_fp128* [[TMP10]], i64 [[TMP3]], i64 [[TMP5]], double* [[VLA]], i64 [[TMP12]]) #[[ATTR4:[0-9]+]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
-// CHECK17-NEXT:    [[TMP62:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP62]])
+// CHECK17-NEXT:    [[TMP70:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP70]])
 // CHECK17-NEXT:    ret void
 //
 //
@@ -2083,15 +2236,32 @@ void array_func(float a[3], St s[2], int n, long double vla1[n]) {
 // CHECK17-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK17-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK17-NEXT:    [[TMP78:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2St7St_funcEPS_iPg_l144.region_id, i32 10, i8** [[TMP75]], i8** [[TMP76]], i64* [[TMP77]], i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK17-NEXT:    [[TMP79:%.*]] = icmp ne i32 [[TMP78]], 0
-// CHECK17-NEXT:    br i1 [[TMP79]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK17-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK17-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK17-NEXT:    store i32 1, i32* [[TMP78]], align 4
+// CHECK17-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK17-NEXT:    store i32 10, i32* [[TMP79]], align 4
+// CHECK17-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK17-NEXT:    store i8** [[TMP75]], i8*** [[TMP80]], align 8
+// CHECK17-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK17-NEXT:    store i8** [[TMP76]], i8*** [[TMP81]], align 8
+// CHECK17-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK17-NEXT:    store i64* [[TMP77]], i64** [[TMP82]], align 8
+// CHECK17-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK17-NEXT:    store i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP83]], align 8
+// CHECK17-NEXT:    [[TMP84:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP84]], align 8
+// CHECK17-NEXT:    [[TMP85:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK17-NEXT:    store i8** null, i8*** [[TMP85]], align 8
+// CHECK17-NEXT:    [[TMP86:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2St7St_funcEPS_iPg_l144.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK17-NEXT:    [[TMP87:%.*]] = icmp ne i32 [[TMP86]], 0
+// CHECK17-NEXT:    br i1 [[TMP87]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK17:       omp_offload.failed:
 // CHECK17-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2St7St_funcEPS_iPg_l144(%struct.St* [[TMP9]], i64 [[TMP1]], ppc_fp128* [[TMP10]], i64 [[TMP3]], i64 [[TMP5]], double* [[VLA]], %struct.St* [[THIS1]], i64 [[TMP12]]) #[[ATTR4]]
 // CHECK17-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK17:       omp_offload.cont:
-// CHECK17-NEXT:    [[TMP80:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
-// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP80]])
+// CHECK17-NEXT:    [[TMP88:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
+// CHECK17-NEXT:    call void @llvm.stackrestore(i8* [[TMP88]])
 // CHECK17-NEXT:    ret void
 //
 //
@@ -2308,15 +2478,32 @@ void array_func(float a[3], St s[2], int n, long double vla1[n]) {
 // CHECK19-NEXT:    [[TMP55:%.*]] = getelementptr inbounds [8 x i8*], [8 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP56:%.*]] = getelementptr inbounds [8 x i8*], [8 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP57:%.*]] = getelementptr inbounds [8 x i64], [8 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP58:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z10array_funcPfP2StiPe_l152.region_id, i32 8, i8** [[TMP55]], i8** [[TMP56]], i64* [[TMP57]], i64* getelementptr inbounds ([8 x i64], [8 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP59:%.*]] = icmp ne i32 [[TMP58]], 0
-// CHECK19-NEXT:    br i1 [[TMP59]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP58:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP58]], align 4
+// CHECK19-NEXT:    [[TMP59:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 8, i32* [[TMP59]], align 4
+// CHECK19-NEXT:    [[TMP60:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP55]], i8*** [[TMP60]], align 4
+// CHECK19-NEXT:    [[TMP61:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP56]], i8*** [[TMP61]], align 4
+// CHECK19-NEXT:    [[TMP62:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP57]], i64** [[TMP62]], align 4
+// CHECK19-NEXT:    [[TMP63:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([8 x i64], [8 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP63]], align 4
+// CHECK19-NEXT:    [[TMP64:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP64]], align 4
+// CHECK19-NEXT:    [[TMP65:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP65]], align 4
+// CHECK19-NEXT:    [[TMP66:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z10array_funcPfP2StiPe_l152.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP67:%.*]] = icmp ne i32 [[TMP66]], 0
+// CHECK19-NEXT:    br i1 [[TMP67]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z10array_funcPfP2StiPe_l152(float* [[TMP5]], %struct.St* [[TMP6]], i32 [[TMP0]], x86_fp80* [[TMP7]], i32 [[TMP1]], i32 [[TMP2]], double* [[VLA]], i32 [[TMP9]]) #[[ATTR4:[0-9]+]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
-// CHECK19-NEXT:    [[TMP60:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP60]])
+// CHECK19-NEXT:    [[TMP68:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP68]])
 // CHECK19-NEXT:    ret void
 //
 //
@@ -2542,15 +2729,32 @@ void array_func(float a[3], St s[2], int n, long double vla1[n]) {
 // CHECK19-NEXT:    [[TMP73:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP74:%.*]] = getelementptr inbounds [10 x i8*], [10 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
 // CHECK19-NEXT:    [[TMP75:%.*]] = getelementptr inbounds [10 x i64], [10 x i64]* [[DOTOFFLOAD_SIZES]], i32 0, i32 0
-// CHECK19-NEXT:    [[TMP76:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2St7St_funcEPS_iPe_l144.region_id, i32 10, i8** [[TMP73]], i8** [[TMP74]], i64* [[TMP75]], i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.3, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK19-NEXT:    [[TMP77:%.*]] = icmp ne i32 [[TMP76]], 0
-// CHECK19-NEXT:    br i1 [[TMP77]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK19-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK19-NEXT:    [[TMP76:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK19-NEXT:    store i32 1, i32* [[TMP76]], align 4
+// CHECK19-NEXT:    [[TMP77:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK19-NEXT:    store i32 10, i32* [[TMP77]], align 4
+// CHECK19-NEXT:    [[TMP78:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK19-NEXT:    store i8** [[TMP73]], i8*** [[TMP78]], align 4
+// CHECK19-NEXT:    [[TMP79:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK19-NEXT:    store i8** [[TMP74]], i8*** [[TMP79]], align 4
+// CHECK19-NEXT:    [[TMP80:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK19-NEXT:    store i64* [[TMP75]], i64** [[TMP80]], align 4
+// CHECK19-NEXT:    [[TMP81:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK19-NEXT:    store i64* getelementptr inbounds ([10 x i64], [10 x i64]* @.offload_maptypes.3, i32 0, i32 0), i64** [[TMP81]], align 4
+// CHECK19-NEXT:    [[TMP82:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP82]], align 4
+// CHECK19-NEXT:    [[TMP83:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK19-NEXT:    store i8** null, i8*** [[TMP83]], align 4
+// CHECK19-NEXT:    [[TMP84:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2St7St_funcEPS_iPe_l144.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK19-NEXT:    [[TMP85:%.*]] = icmp ne i32 [[TMP84]], 0
+// CHECK19-NEXT:    br i1 [[TMP85]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK19:       omp_offload.failed:
 // CHECK19-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2St7St_funcEPS_iPe_l144(%struct.St* [[TMP6]], i32 [[TMP0]], x86_fp80* [[TMP7]], i32 [[TMP1]], i32 [[TMP2]], double* [[VLA]], %struct.St* [[THIS1]], i32 [[TMP9]]) #[[ATTR4]]
 // CHECK19-NEXT:    br label [[OMP_OFFLOAD_CONT]]
 // CHECK19:       omp_offload.cont:
-// CHECK19-NEXT:    [[TMP78:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
-// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP78]])
+// CHECK19-NEXT:    [[TMP86:%.*]] = load i8*, i8** [[SAVED_STACK]], align 4
+// CHECK19-NEXT:    call void @llvm.stackrestore(i8* [[TMP86]])
 // CHECK19-NEXT:    ret void
 //
 //

diff --git a/clang/test/OpenMP/teams_private_codegen.cpp b/clang/test/OpenMP/teams_private_codegen.cpp
index e1fa068..898d696 100644 (file)
--- a/clang/test/OpenMP/teams_private_codegen.cpp
+++ b/clang/test/OpenMP/teams_private_codegen.cpp
@@ -219,9 +219,26 @@ int main() {
 // CHECK1-NEXT:    store i8* null, i8** [[TMP5]], align 8
 // CHECK1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK1-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSC1ERi_l48.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK1-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
-// CHECK1-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK1-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK1-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK1-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK1-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK1-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK1-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK1-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK1-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK1-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK1-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSC1ERi_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK1-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
+// CHECK1-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK1:       omp_offload.failed:
 // CHECK1-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSC1ERi_l48(%struct.SS* [[THIS1]]) #[[ATTR4:[0-9]+]]
 // CHECK1-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -387,9 +404,26 @@ int main() {
 // CHECK3-NEXT:    store i8* null, i8** [[TMP5]], align 4
 // CHECK3-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK3-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK3-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSC1ERi_l48.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK3-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
-// CHECK3-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK3-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK3-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK3-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK3-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK3-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK3-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK3-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK3-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK3-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK3-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK3-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK3-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK3-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK3-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK3-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK3-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSC1ERi_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK3-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
+// CHECK3-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK3:       omp_offload.failed:
 // CHECK3-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSC1ERi_l48(%struct.SS* [[THIS1]]) #[[ATTR4:[0-9]+]]
 // CHECK3-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -520,9 +554,26 @@ int main() {
 // CHECK9-NEXT:    [[ARRAYINIT_ELEMENT:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYINIT_BEGIN]], i64 1
 // CHECK9-NEXT:    call void @_ZN1SIfEC1Ef(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYINIT_ELEMENT]], float noundef 2.000000e+00)
 // CHECK9-NEXT:    call void @_ZN1SIfEC1Ef(%struct.S* noundef nonnull align 4 dereferenceable(4) [[VAR]], float noundef 3.000000e+00)
-// CHECK9-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l136.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK9-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK9-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l136.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK9-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l136() #[[ATTR4:[0-9]+]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -531,18 +582,18 @@ int main() {
 // CHECK9-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[VAR]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE1:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done1:
 // CHECK9-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP4]]
+// CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN2SSC1ERi
@@ -663,9 +714,26 @@ int main() {
 // CHECK9-NEXT:    [[ARRAYINIT_ELEMENT:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYINIT_BEGIN]], i64 1
 // CHECK9-NEXT:    call void @_ZN1SIiEC1Ei(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYINIT_ELEMENT]], i32 noundef signext 2)
 // CHECK9-NEXT:    call void @_ZN1SIiEC1Ei(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[VAR]], i32 noundef signext 3)
-// CHECK9-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l86.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK9-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK9-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP3]], align 8
+// CHECK9-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP4]], align 8
+// CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* null, i64** [[TMP5]], align 8
+// CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* null, i64** [[TMP6]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP7]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP8]], align 8
+// CHECK9-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l86.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK9-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l86() #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -673,18 +741,18 @@ int main() {
 // CHECK9-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[VAR]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i64 2
 // CHECK9-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK9:       arraydestroy.body:
-// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK9-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK9-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK9-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE1:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK9:       arraydestroy.done1:
 // CHECK9-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK9-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK9-NEXT:    ret i32 [[TMP4]]
+// CHECK9-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK9-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK9-LABEL: define {{[^@]+}}@_ZN2SSC2ERi
@@ -718,9 +786,26 @@ int main() {
 // CHECK9-NEXT:    store i8* null, i8** [[TMP5]], align 8
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSC1ERi_l48.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
-// CHECK9-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP15]], align 8
+// CHECK9-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSC1ERi_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
+// CHECK9-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSC1ERi_l48(%struct.SS* [[THIS1]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -928,9 +1013,26 @@ int main() {
 // CHECK9-NEXT:    store i8* null, i8** [[TMP4]], align 8
 // CHECK9-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK9-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK9-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN3SSTIiEC1Ev_l64.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK9-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK9-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK9-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK9-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK9-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK9-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK9-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK9-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK9-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 8
+// CHECK9-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK9-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 8
+// CHECK9-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP11]], align 8
+// CHECK9-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK9-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP12]], align 8
+// CHECK9-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP13]], align 8
+// CHECK9-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK9-NEXT:    store i8** null, i8*** [[TMP14]], align 8
+// CHECK9-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN3SSTIiEC1Ev_l64.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK9-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK9-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK9:       omp_offload.failed:
 // CHECK9-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN3SSTIiEC1Ev_l64(%struct.SST* [[THIS1]]) #[[ATTR4]]
 // CHECK9-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1019,9 +1121,26 @@ int main() {
 // CHECK11-NEXT:    [[ARRAYINIT_ELEMENT:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYINIT_BEGIN]], i32 1
 // CHECK11-NEXT:    call void @_ZN1SIfEC1Ef(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYINIT_ELEMENT]], float noundef 2.000000e+00)
 // CHECK11-NEXT:    call void @_ZN1SIfEC1Ef(%struct.S* noundef nonnull align 4 dereferenceable(4) [[VAR]], float noundef 3.000000e+00)
-// CHECK11-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l136.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK11-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1:[0-9]+]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l136.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK11-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l136() #[[ATTR4:[0-9]+]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1030,18 +1149,18 @@ int main() {
 // CHECK11-NEXT:    store i32 [[CALL]], i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[VAR]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE1:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done1:
 // CHECK11-NEXT:    call void @_ZN1SIfED1Ev(%struct.S* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP4]]
+// CHECK11-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN2SSC1ERi
@@ -1162,9 +1281,26 @@ int main() {
 // CHECK11-NEXT:    [[ARRAYINIT_ELEMENT:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYINIT_BEGIN]], i32 1
 // CHECK11-NEXT:    call void @_ZN1SIiEC1Ei(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYINIT_ELEMENT]], i32 noundef 2)
 // CHECK11-NEXT:    call void @_ZN1SIiEC1Ei(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[VAR]], i32 noundef 3)
-// CHECK11-NEXT:    [[TMP1:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l86.region_id, i32 0, i8** null, i8** null, i64* null, i64* null, i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0
-// CHECK11-NEXT:    br i1 [[TMP2]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP1]], align 4
+// CHECK11-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 0, i32* [[TMP2]], align 4
+// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP3]], align 4
+// CHECK11-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP4]], align 4
+// CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* null, i64** [[TMP5]], align 4
+// CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* null, i64** [[TMP6]], align 4
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l86.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP10:%.*]] = icmp ne i32 [[TMP9]], 0
+// CHECK11-NEXT:    br i1 [[TMP10]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z5tmainIiET_v_l86() #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1172,18 +1308,18 @@ int main() {
 // CHECK11-NEXT:    store i32 0, i32* [[RETVAL]], align 4
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[VAR]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S.0], [2 x %struct.S.0]* [[S_ARR]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAY_BEGIN]], i32 2
 // CHECK11-NEXT:    br label [[ARRAYDESTROY_BODY:%.*]]
 // CHECK11:       arraydestroy.body:
-// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP3]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
+// CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S.0* [ [[TMP11]], [[OMP_OFFLOAD_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
 // CHECK11-NEXT:    [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S_0]], %struct.S.0* [[ARRAYDESTROY_ELEMENTPAST]], i32 -1
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4]]
 // CHECK11-NEXT:    [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S.0* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN]]
 // CHECK11-NEXT:    br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE1:%.*]], label [[ARRAYDESTROY_BODY]]
 // CHECK11:       arraydestroy.done1:
 // CHECK11-NEXT:    call void @_ZN1SIiED1Ev(%struct.S.0* noundef nonnull align 4 dereferenceable(4) [[TEST]]) #[[ATTR4]]
-// CHECK11-NEXT:    [[TMP4:%.*]] = load i32, i32* [[RETVAL]], align 4
-// CHECK11-NEXT:    ret i32 [[TMP4]]
+// CHECK11-NEXT:    [[TMP12:%.*]] = load i32, i32* [[RETVAL]], align 4
+// CHECK11-NEXT:    ret i32 [[TMP12]]
 //
 //
 // CHECK11-LABEL: define {{[^@]+}}@_ZN2SSC2ERi
@@ -1217,9 +1353,26 @@ int main() {
 // CHECK11-NEXT:    store i8* null, i8** [[TMP5]], align 4
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP8:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSC1ERi_l48.region_id, i32 1, i8** [[TMP6]], i8** [[TMP7]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
-// CHECK11-NEXT:    br i1 [[TMP9]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP7]], i8*** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i64** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP15]], align 4
+// CHECK11-NEXT:    [[TMP16:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSC1ERi_l48.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP16]], 0
+// CHECK11-NEXT:    br i1 [[TMP17]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN2SSC1ERi_l48(%struct.SS* [[THIS1]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]
@@ -1427,9 +1580,26 @@ int main() {
 // CHECK11-NEXT:    store i8* null, i8** [[TMP4]], align 4
 // CHECK11-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
 // CHECK11-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
-// CHECK11-NEXT:    [[TMP7:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN3SSTIiEC1Ev_l64.region_id, i32 1, i8** [[TMP5]], i8** [[TMP6]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i8** null, i8** null, i32 0, i32 0)
-// CHECK11-NEXT:    [[TMP8:%.*]] = icmp ne i32 [[TMP7]], 0
-// CHECK11-NEXT:    br i1 [[TMP8]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
+// CHECK11-NEXT:    [[KERNEL_ARGS:%.*]] = alloca [[STRUCT___TGT_KERNEL_ARGUMENTS:%.*]], align 8
+// CHECK11-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 0
+// CHECK11-NEXT:    store i32 1, i32* [[TMP7]], align 4
+// CHECK11-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 1
+// CHECK11-NEXT:    store i32 1, i32* [[TMP8]], align 4
+// CHECK11-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 2
+// CHECK11-NEXT:    store i8** [[TMP5]], i8*** [[TMP9]], align 4
+// CHECK11-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 3
+// CHECK11-NEXT:    store i8** [[TMP6]], i8*** [[TMP10]], align 4
+// CHECK11-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 4
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes.4, i32 0, i32 0), i64** [[TMP11]], align 4
+// CHECK11-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 5
+// CHECK11-NEXT:    store i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes.5, i32 0, i32 0), i64** [[TMP12]], align 4
+// CHECK11-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 6
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP13]], align 4
+// CHECK11-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TGT_KERNEL_ARGUMENTS]], %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]], i32 0, i32 7
+// CHECK11-NEXT:    store i8** null, i8*** [[TMP14]], align 4
+// CHECK11-NEXT:    [[TMP15:%.*]] = call i32 @__tgt_target_kernel(%struct.ident_t* @[[GLOB1]], i64 -1, i32 0, i32 0, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN3SSTIiEC1Ev_l64.region_id, %struct.__tgt_kernel_arguments* [[KERNEL_ARGS]])
+// CHECK11-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+// CHECK11-NEXT:    br i1 [[TMP16]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
 // CHECK11:       omp_offload.failed:
 // CHECK11-NEXT:    call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__ZN3SSTIiEC1Ev_l64(%struct.SST* [[THIS1]]) #[[ATTR4]]
 // CHECK11-NEXT:    br label [[OMP_OFFLOAD_CONT]]

diff --git a/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h b/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h
index 8a6b1c7..77801ae 100644 (file)
--- a/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h
+++ b/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h
@@ -821,6 +821,23 @@ public:
                                 omp::Directive CanceledDirective,
                                 FinalizeCallbackTy ExitCB = {});
 
+  /// Generate a target region entry call.
+  ///
+  /// \param Loc The location at which the request originated and is fulfilled.
+  /// \param Return Return value of the created function returned by reference.
+  /// \param DeviceID Identifier for the device via the 'device' clause.
+  /// \param NumTeams Numer of teams for the region via the 'num_teams' clause
+  ///                 or 0 if unspecified and -1 if there is no 'teams' clause.
+  /// \param NumThreads Number of threads via the 'thread_limit' clause.
+  /// \param HostPtr Pointer to the host-side pointer of the target kernel.
+  /// \param KernelArgs Array of arguments to the kernel.
+  /// \param NoWaitKernelArgs Optional array of arguments to the nowait kernel.
+  InsertPointTy emitTargetKernel(const LocationDescription &Loc, Value *&Return,
+                                 Value *Ident, Value *DeviceID, Value *NumTeams,
+                                 Value *NumThreads, Value *HostPtr,
+                                 ArrayRef<Value *> KernelArgs,
+                                 ArrayRef<Value *> NoWaitArgs = {});
+
   /// Generate a barrier runtime call.
   ///
   /// \param Loc The location at which the request originated and is fulfilled.

diff --git a/llvm/include/llvm/Frontend/OpenMP/OMPKinds.def b/llvm/include/llvm/Frontend/OpenMP/OMPKinds.def
index 14aa53a..15d1922 100644 (file)
--- a/llvm/include/llvm/Frontend/OpenMP/OMPKinds.def
+++ b/llvm/include/llvm/Frontend/OpenMP/OMPKinds.def
@@ -88,6 +88,8 @@ __OMP_ARRAY_TYPE(KmpCriticalName, Int32, 8)
 __OMP_STRUCT_TYPE(Ident, ident_t, Int32, Int32, Int32, Int32, Int8Ptr)
 __OMP_STRUCT_TYPE(OffloadEntry, __tgt_offload_entry, Int8Ptr, Int8Ptr, SizeTy,
                   Int32, Int32)
+__OMP_STRUCT_TYPE(KernelArgs, __tgt_kernel_arguments, Int32, Int32, VoidPtrPtr,
+                  VoidPtrPtr, Int64Ptr, Int64Ptr, VoidPtrPtr, VoidPtrPtr)
 __OMP_STRUCT_TYPE(AsyncInfo, __tgt_async_info, Int8Ptr)
 
 #undef __OMP_STRUCT_TYPE
@@ -412,6 +414,10 @@ __OMP_RTL(__tgt_target_teams_mapper, false, Int32, IdentPtr, Int64, VoidPtr, Int
 __OMP_RTL(__tgt_target_teams_nowait_mapper, false, Int32, IdentPtr, Int64,
           VoidPtr, Int32, VoidPtrPtr, VoidPtrPtr, Int64Ptr, Int64Ptr,
           VoidPtrPtr, VoidPtrPtr, Int32, Int32, Int32, VoidPtr, Int32, VoidPtr)
+__OMP_RTL(__tgt_target_kernel, false, Int32, IdentPtr, Int64, Int32, Int32,
+          VoidPtr, KernelArgsPtr)
+__OMP_RTL(__tgt_target_kernel_nowait, false, Int32, IdentPtr, Int64, Int32,
+          Int32, VoidPtr, KernelArgsPtr, Int32, VoidPtr, Int32, VoidPtr)
 __OMP_RTL(__tgt_register_requires, false, Void, Int64)
 __OMP_RTL(__tgt_target_data_begin_mapper, false, Void, IdentPtr, Int64, Int32, VoidPtrPtr,
           VoidPtrPtr, Int64Ptr, Int64Ptr, VoidPtrPtr, VoidPtrPtr)
@@ -937,6 +943,10 @@ __OMP_RTL_ATTRS(__tgt_target_teams_mapper, ForkAttrs, AttributeSet(),
                 ParamAttrs())
 __OMP_RTL_ATTRS(__tgt_target_teams_nowait_mapper, ForkAttrs, AttributeSet(),
                 ParamAttrs())
+__OMP_RTL_ATTRS(__tgt_target_kernel, ForkAttrs, AttributeSet(),
+                ParamAttrs())
+__OMP_RTL_ATTRS(__tgt_target_kernel_nowait, ForkAttrs, AttributeSet(),
+                ParamAttrs())
 __OMP_RTL_ATTRS(__tgt_register_requires, ForkAttrs, AttributeSet(),
                 ParamAttrs())
 __OMP_RTL_ATTRS(__tgt_target_data_begin_mapper, ForkAttrs, AttributeSet(),

diff --git a/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp b/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
index a540bfb..f065ecc 100644 (file)
--- a/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
+++ b/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
@@ -791,6 +791,38 @@ void OpenMPIRBuilder::emitOffloadingEntry(Constant *Addr, StringRef Name,
   Entry->setAlignment(Align(1));
 }
 
+OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitTargetKernel(
+    const LocationDescription &Loc, Value *&Return, Value *Ident,
+    Value *DeviceID, Value *NumTeams, Value *NumThreads, Value *HostPtr,
+    ArrayRef<Value *> KernelArgs, ArrayRef<Value *> NoWaitArgs) {
+  if (!updateToLocation(Loc))
+    return Loc.IP;
+
+  auto *KernelArgsPtr =
+      Builder.CreateAlloca(OpenMPIRBuilder::KernelArgs, nullptr, "kernel_args");
+  for (unsigned I = 0, Size = KernelArgs.size(); I != Size; ++I) {
+    llvm::Value *Arg =
+        Builder.CreateStructGEP(OpenMPIRBuilder::KernelArgs, KernelArgsPtr, I);
+    Builder.CreateAlignedStore(
+        KernelArgs[I], Arg,
+        M.getDataLayout().getPrefTypeAlign(KernelArgs[I]->getType()));
+  }
+
+  bool HasNoWait = !NoWaitArgs.empty();
+  SmallVector<Value *> OffloadingArgs{Ident,      DeviceID, NumTeams,
+                                      NumThreads, HostPtr,  KernelArgsPtr};
+  if (HasNoWait)
+    OffloadingArgs.append(NoWaitArgs.begin(), NoWaitArgs.end());
+
+  Return = Builder.CreateCall(
+      HasNoWait
+          ? getOrCreateRuntimeFunction(M, OMPRTL___tgt_target_kernel_nowait)
+          : getOrCreateRuntimeFunction(M, OMPRTL___tgt_target_kernel),
+      OffloadingArgs);
+
+  return Builder.saveIP();
+}
+
 void OpenMPIRBuilder::emitCancelationCheckImpl(Value *CancelFlag,
                                                omp::Directive CanceledDirective,
                                                FinalizeCallbackTy ExitCB) {